blob: a0e0c82c1e4cd22324e87a86a5df7301f35b8a83 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
Tejun Heo88999a82010-04-09 18:57:01 +09002 * mm/percpu.c - percpu memory allocator
Tejun Heofbf59bc2009-02-20 16:29:08 +09003 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
Dennis Zhou (Facebook)5e81ee32017-07-24 19:02:20 -04007 * Copyright (C) 2017 Facebook Inc.
8 * Copyright (C) 2017 Dennis Zhou <dennisszhou@gmail.com>
9 *
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -040010 * This file is released under the GPLv2 license.
Tejun Heofbf59bc2009-02-20 16:29:08 +090011 *
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -040012 * The percpu allocator handles both static and dynamic areas. Percpu
13 * areas are allocated in chunks which are divided into units. There is
14 * a 1-to-1 mapping for units to possible cpus. These units are grouped
15 * based on NUMA properties of the machine.
Tejun Heofbf59bc2009-02-20 16:29:08 +090016 *
17 * c0 c1 c2
18 * ------------------- ------------------- ------------
19 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
20 * ------------------- ...... ------------------- .... ------------
21 *
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -040022 * Allocation is done by offsets into a unit's address space. Ie., an
23 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
24 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
25 * and even sparse. Access is handled by configuring percpu base
26 * registers according to the cpu to unit mappings and offsetting the
27 * base address using pcpu_unit_size.
Tejun Heofbf59bc2009-02-20 16:29:08 +090028 *
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -040029 * There is special consideration for the first chunk which must handle
30 * the static percpu variables in the kernel image as allocation services
Dennis Zhou (Facebook)5e81ee32017-07-24 19:02:20 -040031 * are not online yet. In short, the first chunk is structured like so:
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -040032 *
33 * <Static | [Reserved] | Dynamic>
34 *
35 * The static data is copied from the original section managed by the
36 * linker. The reserved section, if non-zero, primarily manages static
37 * percpu variables from kernel modules. Finally, the dynamic section
38 * takes care of normal allocations.
Tejun Heofbf59bc2009-02-20 16:29:08 +090039 *
Dennis Zhou (Facebook)5e81ee32017-07-24 19:02:20 -040040 * The allocator organizes chunks into lists according to free size and
41 * tries to allocate from the fullest chunk first. Each chunk is managed
42 * by a bitmap with metadata blocks. The allocation map is updated on
43 * every allocation and free to reflect the current state while the boundary
44 * map is only updated on allocation. Each metadata block contains
45 * information to help mitigate the need to iterate over large portions
46 * of the bitmap. The reverse mapping from page to chunk is stored in
47 * the page's index. Lastly, units are lazily backed and grow in unison.
Tejun Heofbf59bc2009-02-20 16:29:08 +090048 *
Dennis Zhou (Facebook)5e81ee32017-07-24 19:02:20 -040049 * There is a unique conversion that goes on here between bytes and bits.
50 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
51 * tracks the number of pages it is responsible for in nr_pages. Helper
52 * functions are used to convert from between the bytes, bits, and blocks.
53 * All hints are managed in bits unless explicitly stated.
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -040054 *
Masahiro Yamada4091fb92017-02-27 14:29:56 -080055 * To use this allocator, arch code should do the following:
Tejun Heofbf59bc2009-02-20 16:29:08 +090056 *
Tejun Heofbf59bc2009-02-20 16:29:08 +090057 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090058 * regular address to percpu pointer and back if they need to be
59 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090060 *
Tejun Heo8d408b42009-02-24 11:57:21 +090061 * - use pcpu_setup_first_chunk() during percpu area initialization to
62 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090063 */
64
Joe Perches870d4b12016-03-17 14:19:53 -070065#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
66
Tejun Heofbf59bc2009-02-20 16:29:08 +090067#include <linux/bitmap.h>
68#include <linux/bootmem.h>
Tejun Heofd1e8a12009-08-14 15:00:51 +090069#include <linux/err.h>
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -040070#include <linux/lcm.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090071#include <linux/list.h>
Tejun Heoa530b792009-07-04 08:11:00 +090072#include <linux/log2.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090073#include <linux/mm.h>
74#include <linux/module.h>
75#include <linux/mutex.h>
76#include <linux/percpu.h>
77#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090078#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090079#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090080#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090081#include <linux/workqueue.h>
Catalin Marinasf528f0b2011-09-26 17:12:53 +010082#include <linux/kmemleak.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090083
84#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090085#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090086#include <asm/tlbflush.h>
Vivek Goyal3b034b02009-11-24 15:50:03 +090087#include <asm/io.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090088
Dennis Zhoudf95e792017-06-19 19:28:32 -040089#define CREATE_TRACE_POINTS
90#include <trace/events/percpu.h>
91
Dennis Zhou8fa3ed82017-06-19 19:28:30 -040092#include "percpu-internal.h"
93
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -070094/* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
95#define PCPU_SLOT_BASE_SHIFT 5
96
Tejun Heo1a4d7602014-09-02 14:46:05 -040097#define PCPU_EMPTY_POP_PAGES_LOW 2
98#define PCPU_EMPTY_POP_PAGES_HIGH 4
Tejun Heofbf59bc2009-02-20 16:29:08 +090099
Tejun Heobbddff02010-09-03 18:22:48 +0200100#ifdef CONFIG_SMP
Tejun Heoe0100982009-03-10 16:27:48 +0900101/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
102#ifndef __addr_to_pcpu_ptr
103#define __addr_to_pcpu_ptr(addr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +0900104 (void __percpu *)((unsigned long)(addr) - \
105 (unsigned long)pcpu_base_addr + \
106 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +0900107#endif
108#ifndef __pcpu_ptr_to_addr
109#define __pcpu_ptr_to_addr(ptr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +0900110 (void __force *)((unsigned long)(ptr) + \
111 (unsigned long)pcpu_base_addr - \
112 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +0900113#endif
Tejun Heobbddff02010-09-03 18:22:48 +0200114#else /* CONFIG_SMP */
115/* on UP, it's always identity mapped */
116#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
117#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
118#endif /* CONFIG_SMP */
Tejun Heoe0100982009-03-10 16:27:48 +0900119
Daniel Micay13287102017-05-10 13:36:37 -0400120static int pcpu_unit_pages __ro_after_init;
121static int pcpu_unit_size __ro_after_init;
122static int pcpu_nr_units __ro_after_init;
123static int pcpu_atom_size __ro_after_init;
Dennis Zhou8fa3ed82017-06-19 19:28:30 -0400124int pcpu_nr_slots __ro_after_init;
Daniel Micay13287102017-05-10 13:36:37 -0400125static size_t pcpu_chunk_struct_size __ro_after_init;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900126
Tejun Heoa855b842011-11-18 10:55:35 -0800127/* cpus with the lowest and highest unit addresses */
Daniel Micay13287102017-05-10 13:36:37 -0400128static unsigned int pcpu_low_unit_cpu __ro_after_init;
129static unsigned int pcpu_high_unit_cpu __ro_after_init;
Tejun Heo2f39e632009-07-04 08:11:00 +0900130
Tejun Heofbf59bc2009-02-20 16:29:08 +0900131/* the address of the first chunk which starts with the kernel static area */
Daniel Micay13287102017-05-10 13:36:37 -0400132void *pcpu_base_addr __ro_after_init;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900133EXPORT_SYMBOL_GPL(pcpu_base_addr);
134
Daniel Micay13287102017-05-10 13:36:37 -0400135static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
136const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900137
Tejun Heo65632972009-08-14 15:00:52 +0900138/* group information, used for vm allocation */
Daniel Micay13287102017-05-10 13:36:37 -0400139static int pcpu_nr_groups __ro_after_init;
140static const unsigned long *pcpu_group_offsets __ro_after_init;
141static const size_t *pcpu_group_sizes __ro_after_init;
Tejun Heo65632972009-08-14 15:00:52 +0900142
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900143/*
144 * The first chunk which always exists. Note that unlike other
145 * chunks, this one can be allocated and mapped in several different
146 * ways and thus often doesn't live in the vmalloc area.
147 */
Dennis Zhou8fa3ed82017-06-19 19:28:30 -0400148struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900149
150/*
151 * Optional reserved chunk. This chunk reserves part of the first
Dennis Zhou (Facebook)e2266702017-07-24 19:01:59 -0400152 * chunk and serves it for reserved allocations. When the reserved
153 * region doesn't exist, the following variable is NULL.
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900154 */
Dennis Zhou8fa3ed82017-06-19 19:28:30 -0400155struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
Tejun Heoedcb4632009-03-06 14:33:59 +0900156
Dennis Zhou8fa3ed82017-06-19 19:28:30 -0400157DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
Tejun Heo6710e592016-05-25 11:48:25 -0400158static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900159
Dennis Zhou8fa3ed82017-06-19 19:28:30 -0400160struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900161
Tejun Heo4f996e22016-05-25 11:48:25 -0400162/* chunks which need their map areas extended, protected by pcpu_lock */
163static LIST_HEAD(pcpu_map_extend_chunks);
164
Tejun Heob539b872014-09-02 14:46:05 -0400165/*
166 * The number of empty populated pages, protected by pcpu_lock. The
167 * reserved chunk doesn't contribute to the count.
168 */
Dennis Zhou (Facebook)6b9b6f32017-07-15 22:23:08 -0400169int pcpu_nr_empty_pop_pages;
Tejun Heob539b872014-09-02 14:46:05 -0400170
Tejun Heo1a4d7602014-09-02 14:46:05 -0400171/*
172 * Balance work is used to populate or destroy chunks asynchronously. We
173 * try to keep the number of populated free pages between
174 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
175 * empty chunk.
176 */
Tejun Heofe6bd8c2014-09-02 14:46:05 -0400177static void pcpu_balance_workfn(struct work_struct *work);
178static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
Tejun Heo1a4d7602014-09-02 14:46:05 -0400179static bool pcpu_async_enabled __read_mostly;
180static bool pcpu_atomic_alloc_failed;
181
182static void pcpu_schedule_balance_work(void)
183{
184 if (pcpu_async_enabled)
185 schedule_work(&pcpu_balance_work);
186}
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900187
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400188/**
Dennis Zhou (Facebook)560f2c22017-07-24 19:02:06 -0400189 * pcpu_addr_in_chunk - check if the address is served from this chunk
190 * @chunk: chunk of interest
191 * @addr: percpu address
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400192 *
193 * RETURNS:
Dennis Zhou (Facebook)560f2c22017-07-24 19:02:06 -0400194 * True if the address is served from this chunk.
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400195 */
Dennis Zhou (Facebook)560f2c22017-07-24 19:02:06 -0400196static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
Tejun Heo020ec652010-04-09 18:57:00 +0900197{
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400198 void *start_addr, *end_addr;
Tejun Heo020ec652010-04-09 18:57:00 +0900199
Dennis Zhou (Facebook)560f2c22017-07-24 19:02:06 -0400200 if (!chunk)
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400201 return false;
202
Dennis Zhou (Facebook)560f2c22017-07-24 19:02:06 -0400203 start_addr = chunk->base_addr + chunk->start_offset;
204 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
205 chunk->end_offset;
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400206
207 return addr >= start_addr && addr < end_addr;
Tejun Heo020ec652010-04-09 18:57:00 +0900208}
209
Tejun Heod9b55ee2009-02-24 11:57:21 +0900210static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900211{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900212 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900213 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
214}
215
Tejun Heod9b55ee2009-02-24 11:57:21 +0900216static int pcpu_size_to_slot(int size)
217{
218 if (size == pcpu_unit_size)
219 return pcpu_nr_slots - 1;
220 return __pcpu_size_to_slot(size);
221}
222
Tejun Heofbf59bc2009-02-20 16:29:08 +0900223static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
224{
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700225 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900226 return 0;
227
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700228 return pcpu_size_to_slot(chunk->free_bytes);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900229}
230
Tejun Heo88999a82010-04-09 18:57:01 +0900231/* set the pointer to a chunk in a page struct */
232static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
233{
234 page->index = (unsigned long)pcpu;
235}
236
237/* obtain pointer to a chunk from a page struct */
238static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
239{
240 return (struct pcpu_chunk *)page->index;
241}
242
243static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900244{
Tejun Heo2f39e632009-07-04 08:11:00 +0900245 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900246}
247
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400248static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
249{
250 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
251}
252
Tejun Heo9983b6f02010-06-18 11:44:31 +0200253static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
254 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900255{
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -0400256 return (unsigned long)chunk->base_addr +
257 pcpu_unit_page_offset(cpu, page_idx);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900258}
259
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -0400260static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900261{
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -0400262 *rs = find_next_zero_bit(bitmap, end, *rs);
263 *re = find_next_bit(bitmap, end, *rs + 1);
Tejun Heoce3141a2009-07-04 08:11:00 +0900264}
265
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -0400266static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900267{
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -0400268 *rs = find_next_bit(bitmap, end, *rs);
269 *re = find_next_zero_bit(bitmap, end, *rs + 1);
Tejun Heoce3141a2009-07-04 08:11:00 +0900270}
271
272/*
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -0400273 * Bitmap region iterators. Iterates over the bitmap between
274 * [@start, @end) in @chunk. @rs and @re should be integer variables
275 * and will be set to start and end index of the current free region.
Tejun Heoce3141a2009-07-04 08:11:00 +0900276 */
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -0400277#define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \
278 for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
279 (rs) < (re); \
280 (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
Tejun Heoce3141a2009-07-04 08:11:00 +0900281
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -0400282#define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \
283 for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \
284 (rs) < (re); \
285 (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
Tejun Heoce3141a2009-07-04 08:11:00 +0900286
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400287/*
288 * The following are helper functions to help access bitmaps and convert
289 * between bitmap offsets to address offsets.
290 */
291static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
292{
293 return chunk->alloc_map +
294 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
295}
296
297static unsigned long pcpu_off_to_block_index(int off)
298{
299 return off / PCPU_BITMAP_BLOCK_BITS;
300}
301
302static unsigned long pcpu_off_to_block_off(int off)
303{
304 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
305}
306
Dennis Zhou (Facebook)b185cd02017-07-24 19:02:17 -0400307static unsigned long pcpu_block_off_to_off(int index, int off)
308{
309 return index * PCPU_BITMAP_BLOCK_BITS + off;
310}
311
Tejun Heofbf59bc2009-02-20 16:29:08 +0900312/**
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400313 * pcpu_next_md_free_region - finds the next hint free area
314 * @chunk: chunk of interest
315 * @bit_off: chunk offset
316 * @bits: size of free area
317 *
318 * Helper function for pcpu_for_each_md_free_region. It checks
319 * block->contig_hint and performs aggregation across blocks to find the
320 * next hint. It modifies bit_off and bits in-place to be consumed in the
321 * loop.
322 */
323static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
324 int *bits)
325{
326 int i = pcpu_off_to_block_index(*bit_off);
327 int block_off = pcpu_off_to_block_off(*bit_off);
328 struct pcpu_block_md *block;
329
330 *bits = 0;
331 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
332 block++, i++) {
333 /* handles contig area across blocks */
334 if (*bits) {
335 *bits += block->left_free;
336 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
337 continue;
338 return;
339 }
340
341 /*
342 * This checks three things. First is there a contig_hint to
343 * check. Second, have we checked this hint before by
344 * comparing the block_off. Third, is this the same as the
345 * right contig hint. In the last case, it spills over into
346 * the next block and should be handled by the contig area
347 * across blocks code.
348 */
349 *bits = block->contig_hint;
350 if (*bits && block->contig_hint_start >= block_off &&
351 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
352 *bit_off = pcpu_block_off_to_off(i,
353 block->contig_hint_start);
354 return;
355 }
Dennis Zhou1fa4df32017-09-27 16:35:00 -0500356 /* reset to satisfy the second predicate above */
357 block_off = 0;
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400358
359 *bits = block->right_free;
360 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
361 }
362}
363
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400364/**
365 * pcpu_next_fit_region - finds fit areas for a given allocation request
366 * @chunk: chunk of interest
367 * @alloc_bits: size of allocation
368 * @align: alignment of area (max PAGE_SIZE)
369 * @bit_off: chunk offset
370 * @bits: size of free area
371 *
372 * Finds the next free region that is viable for use with a given size and
373 * alignment. This only returns if there is a valid area to be used for this
374 * allocation. block->first_free is returned if the allocation request fits
375 * within the block to see if the request can be fulfilled prior to the contig
376 * hint.
377 */
378static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
379 int align, int *bit_off, int *bits)
380{
381 int i = pcpu_off_to_block_index(*bit_off);
382 int block_off = pcpu_off_to_block_off(*bit_off);
383 struct pcpu_block_md *block;
384
385 *bits = 0;
386 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
387 block++, i++) {
388 /* handles contig area across blocks */
389 if (*bits) {
390 *bits += block->left_free;
391 if (*bits >= alloc_bits)
392 return;
393 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
394 continue;
395 }
396
397 /* check block->contig_hint */
398 *bits = ALIGN(block->contig_hint_start, align) -
399 block->contig_hint_start;
400 /*
401 * This uses the block offset to determine if this has been
402 * checked in the prior iteration.
403 */
404 if (block->contig_hint &&
405 block->contig_hint_start >= block_off &&
406 block->contig_hint >= *bits + alloc_bits) {
407 *bits += alloc_bits + block->contig_hint_start -
408 block->first_free;
409 *bit_off = pcpu_block_off_to_off(i, block->first_free);
410 return;
411 }
Dennis Zhou1fa4df32017-09-27 16:35:00 -0500412 /* reset to satisfy the second predicate above */
413 block_off = 0;
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400414
415 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
416 align);
417 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
418 *bit_off = pcpu_block_off_to_off(i, *bit_off);
419 if (*bits >= alloc_bits)
420 return;
421 }
422
423 /* no valid offsets were found - fail condition */
424 *bit_off = pcpu_chunk_map_bits(chunk);
425}
426
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400427/*
428 * Metadata free area iterators. These perform aggregation of free areas
429 * based on the metadata blocks and return the offset @bit_off and size in
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400430 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
431 * a fit is found for the allocation request.
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400432 */
433#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
434 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
435 (bit_off) < pcpu_chunk_map_bits((chunk)); \
436 (bit_off) += (bits) + 1, \
437 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
438
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400439#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
440 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
441 &(bits)); \
442 (bit_off) < pcpu_chunk_map_bits((chunk)); \
443 (bit_off) += (bits), \
444 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
445 &(bits)))
446
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400447/**
Bob Liu90459ce02011-08-04 11:02:33 +0200448 * pcpu_mem_zalloc - allocate memory
Tejun Heo1880d932009-03-07 00:44:09 +0900449 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900450 *
Tejun Heo1880d932009-03-07 00:44:09 +0900451 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
Bob Liu90459ce02011-08-04 11:02:33 +0200452 * kzalloc() is used; otherwise, vzalloc() is used. The returned
Tejun Heo1880d932009-03-07 00:44:09 +0900453 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900454 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900455 * CONTEXT:
456 * Does GFP_KERNEL allocation.
457 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900458 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900459 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900460 */
Bob Liu90459ce02011-08-04 11:02:33 +0200461static void *pcpu_mem_zalloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900462{
Tejun Heo099a19d2010-06-27 18:50:00 +0200463 if (WARN_ON_ONCE(!slab_is_available()))
464 return NULL;
465
Tejun Heofbf59bc2009-02-20 16:29:08 +0900466 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900467 return kzalloc(size, GFP_KERNEL);
Jesper Juhl7af4c092010-10-30 15:56:54 +0200468 else
469 return vzalloc(size);
Tejun Heo1880d932009-03-07 00:44:09 +0900470}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900471
Tejun Heo1880d932009-03-07 00:44:09 +0900472/**
473 * pcpu_mem_free - free memory
474 * @ptr: memory to free
Tejun Heo1880d932009-03-07 00:44:09 +0900475 *
Bob Liu90459ce02011-08-04 11:02:33 +0200476 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
Tejun Heo1880d932009-03-07 00:44:09 +0900477 */
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -0800478static void pcpu_mem_free(void *ptr)
Tejun Heo1880d932009-03-07 00:44:09 +0900479{
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -0800480 kvfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900481}
482
483/**
484 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
485 * @chunk: chunk of interest
486 * @oslot: the previous slot it was on
487 *
488 * This function is called after an allocation or free changed @chunk.
489 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900490 * moved to the slot. Note that the reserved chunk is never put on
491 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900492 *
493 * CONTEXT:
494 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900495 */
496static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
497{
498 int nslot = pcpu_chunk_slot(chunk);
499
Tejun Heoedcb4632009-03-06 14:33:59 +0900500 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900501 if (oslot < nslot)
502 list_move(&chunk->list, &pcpu_slot[nslot]);
503 else
504 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
505 }
506}
507
Tejun Heofbf59bc2009-02-20 16:29:08 +0900508/**
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700509 * pcpu_cnt_pop_pages- counts populated backing pages in range
Tejun Heo833af842009-11-11 15:35:18 +0900510 * @chunk: chunk of interest
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700511 * @bit_off: start offset
512 * @bits: size of area to check
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900513 *
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700514 * Calculates the number of populated pages in the region
515 * [page_start, page_end). This keeps track of how many empty populated
516 * pages are available and decide if async work should be scheduled.
Tejun Heoccea34b2009-03-07 00:44:13 +0900517 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900518 * RETURNS:
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700519 * The nr of populated pages.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900520 */
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700521static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off,
522 int bits)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900523{
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700524 int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE);
525 int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
526
527 if (page_start >= page_end)
528 return 0;
529
530 /*
531 * bitmap_weight counts the number of bits set in a bitmap up to
532 * the specified number of bits. This is counting the populated
533 * pages up to page_end and then subtracting the populated pages
534 * up to page_start to count the populated pages in
535 * [page_start, page_end).
536 */
537 return bitmap_weight(chunk->populated, page_end) -
538 bitmap_weight(chunk->populated, page_start);
539}
540
541/**
542 * pcpu_chunk_update - updates the chunk metadata given a free area
543 * @chunk: chunk of interest
544 * @bit_off: chunk offset
545 * @bits: size of free area
546 *
Dennis Zhou (Facebook)13f96632017-07-24 19:02:14 -0400547 * This updates the chunk's contig hint and starting offset given a free area.
Dennis Zhou (Facebook)268625a2017-07-24 19:02:15 -0400548 * Choose the best starting offset if the contig hint is equal.
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700549 */
550static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits)
551{
Dennis Zhou (Facebook)13f96632017-07-24 19:02:14 -0400552 if (bits > chunk->contig_bits) {
553 chunk->contig_bits_start = bit_off;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700554 chunk->contig_bits = bits;
Dennis Zhou (Facebook)268625a2017-07-24 19:02:15 -0400555 } else if (bits == chunk->contig_bits && chunk->contig_bits_start &&
556 (!bit_off ||
557 __ffs(bit_off) > __ffs(chunk->contig_bits_start))) {
558 /* use the start with the best alignment */
559 chunk->contig_bits_start = bit_off;
Dennis Zhou (Facebook)13f96632017-07-24 19:02:14 -0400560 }
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700561}
562
563/**
564 * pcpu_chunk_refresh_hint - updates metadata about a chunk
565 * @chunk: chunk of interest
566 *
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400567 * Iterates over the metadata blocks to find the largest contig area.
568 * It also counts the populated pages and uses the delta to update the
569 * global count.
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700570 *
571 * Updates:
572 * chunk->contig_bits
Dennis Zhou (Facebook)13f96632017-07-24 19:02:14 -0400573 * chunk->contig_bits_start
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400574 * nr_empty_pop_pages (chunk and global)
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700575 */
576static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk)
577{
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400578 int bit_off, bits, nr_empty_pop_pages;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700579
580 /* clear metadata */
581 chunk->contig_bits = 0;
582
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400583 bit_off = chunk->first_bit;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700584 bits = nr_empty_pop_pages = 0;
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400585 pcpu_for_each_md_free_region(chunk, bit_off, bits) {
586 pcpu_chunk_update(chunk, bit_off, bits);
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700587
Dennis Zhou (Facebook)525ca842017-07-24 19:02:18 -0400588 nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, bit_off, bits);
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700589 }
590
591 /*
592 * Keep track of nr_empty_pop_pages.
593 *
594 * The chunk maintains the previous number of free pages it held,
595 * so the delta is used to update the global counter. The reserved
596 * chunk is not part of the free page count as they are populated
597 * at init and are special to serving reserved allocations.
598 */
599 if (chunk != pcpu_reserved_chunk)
600 pcpu_nr_empty_pop_pages +=
601 (nr_empty_pop_pages - chunk->nr_empty_pop_pages);
602
603 chunk->nr_empty_pop_pages = nr_empty_pop_pages;
604}
605
606/**
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400607 * pcpu_block_update - updates a block given a free area
608 * @block: block of interest
609 * @start: start offset in block
610 * @end: end offset in block
611 *
612 * Updates a block given a known free area. The region [start, end) is
Dennis Zhou (Facebook)268625a2017-07-24 19:02:15 -0400613 * expected to be the entirety of the free area within a block. Chooses
614 * the best starting offset if the contig hints are equal.
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400615 */
616static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
617{
618 int contig = end - start;
619
620 block->first_free = min(block->first_free, start);
621 if (start == 0)
622 block->left_free = contig;
623
624 if (end == PCPU_BITMAP_BLOCK_BITS)
625 block->right_free = contig;
626
627 if (contig > block->contig_hint) {
628 block->contig_hint_start = start;
629 block->contig_hint = contig;
Dennis Zhou (Facebook)268625a2017-07-24 19:02:15 -0400630 } else if (block->contig_hint_start && contig == block->contig_hint &&
631 (!start || __ffs(start) > __ffs(block->contig_hint_start))) {
632 /* use the start with the best alignment */
633 block->contig_hint_start = start;
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400634 }
635}
636
637/**
638 * pcpu_block_refresh_hint
639 * @chunk: chunk of interest
640 * @index: index of the metadata block
641 *
642 * Scans over the block beginning at first_free and updates the block
643 * metadata accordingly.
644 */
645static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
646{
647 struct pcpu_block_md *block = chunk->md_blocks + index;
648 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
649 int rs, re; /* region start, region end */
650
651 /* clear hints */
652 block->contig_hint = 0;
653 block->left_free = block->right_free = 0;
654
655 /* iterate over free areas and update the contig hints */
656 pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free,
657 PCPU_BITMAP_BLOCK_BITS) {
658 pcpu_block_update(block, rs, re);
659 }
660}
661
662/**
663 * pcpu_block_update_hint_alloc - update hint on allocation path
664 * @chunk: chunk of interest
665 * @bit_off: chunk offset
666 * @bits: size of request
Dennis Zhou (Facebook)fc304332017-07-24 19:02:16 -0400667 *
668 * Updates metadata for the allocation path. The metadata only has to be
669 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
670 * scans are required if the block's contig hint is broken.
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400671 */
672static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
673 int bits)
674{
675 struct pcpu_block_md *s_block, *e_block, *block;
676 int s_index, e_index; /* block indexes of the freed allocation */
677 int s_off, e_off; /* block offsets of the freed allocation */
678
679 /*
680 * Calculate per block offsets.
681 * The calculation uses an inclusive range, but the resulting offsets
682 * are [start, end). e_index always points to the last block in the
683 * range.
684 */
685 s_index = pcpu_off_to_block_index(bit_off);
686 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
687 s_off = pcpu_off_to_block_off(bit_off);
688 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
689
690 s_block = chunk->md_blocks + s_index;
691 e_block = chunk->md_blocks + e_index;
692
693 /*
694 * Update s_block.
Dennis Zhou (Facebook)fc304332017-07-24 19:02:16 -0400695 * block->first_free must be updated if the allocation takes its place.
696 * If the allocation breaks the contig_hint, a scan is required to
697 * restore this hint.
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400698 */
Dennis Zhou (Facebook)fc304332017-07-24 19:02:16 -0400699 if (s_off == s_block->first_free)
700 s_block->first_free = find_next_zero_bit(
701 pcpu_index_alloc_map(chunk, s_index),
702 PCPU_BITMAP_BLOCK_BITS,
703 s_off + bits);
704
705 if (s_off >= s_block->contig_hint_start &&
706 s_off < s_block->contig_hint_start + s_block->contig_hint) {
707 /* block contig hint is broken - scan to fix it */
708 pcpu_block_refresh_hint(chunk, s_index);
709 } else {
710 /* update left and right contig manually */
711 s_block->left_free = min(s_block->left_free, s_off);
712 if (s_index == e_index)
713 s_block->right_free = min_t(int, s_block->right_free,
714 PCPU_BITMAP_BLOCK_BITS - e_off);
715 else
716 s_block->right_free = 0;
717 }
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400718
719 /*
720 * Update e_block.
721 */
722 if (s_index != e_index) {
Dennis Zhou (Facebook)fc304332017-07-24 19:02:16 -0400723 /*
724 * When the allocation is across blocks, the end is along
725 * the left part of the e_block.
726 */
727 e_block->first_free = find_next_zero_bit(
728 pcpu_index_alloc_map(chunk, e_index),
729 PCPU_BITMAP_BLOCK_BITS, e_off);
730
731 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
732 /* reset the block */
733 e_block++;
734 } else {
735 if (e_off > e_block->contig_hint_start) {
736 /* contig hint is broken - scan to fix it */
737 pcpu_block_refresh_hint(chunk, e_index);
738 } else {
739 e_block->left_free = 0;
740 e_block->right_free =
741 min_t(int, e_block->right_free,
742 PCPU_BITMAP_BLOCK_BITS - e_off);
743 }
744 }
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400745
746 /* update in-between md_blocks */
747 for (block = s_block + 1; block < e_block; block++) {
748 block->contig_hint = 0;
749 block->left_free = 0;
750 block->right_free = 0;
751 }
752 }
753
Dennis Zhou (Facebook)fc304332017-07-24 19:02:16 -0400754 /*
755 * The only time a full chunk scan is required is if the chunk
756 * contig hint is broken. Otherwise, it means a smaller space
757 * was used and therefore the chunk contig hint is still correct.
758 */
759 if (bit_off >= chunk->contig_bits_start &&
760 bit_off < chunk->contig_bits_start + chunk->contig_bits)
761 pcpu_chunk_refresh_hint(chunk);
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400762}
763
764/**
765 * pcpu_block_update_hint_free - updates the block hints on the free path
766 * @chunk: chunk of interest
767 * @bit_off: chunk offset
768 * @bits: size of request
Dennis Zhou (Facebook)b185cd02017-07-24 19:02:17 -0400769 *
770 * Updates metadata for the allocation path. This avoids a blind block
771 * refresh by making use of the block contig hints. If this fails, it scans
772 * forward and backward to determine the extent of the free area. This is
773 * capped at the boundary of blocks.
774 *
775 * A chunk update is triggered if a page becomes free, a block becomes free,
776 * or the free spans across blocks. This tradeoff is to minimize iterating
777 * over the block metadata to update chunk->contig_bits. chunk->contig_bits
778 * may be off by up to a page, but it will never be more than the available
779 * space. If the contig hint is contained in one block, it will be accurate.
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400780 */
781static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
782 int bits)
783{
784 struct pcpu_block_md *s_block, *e_block, *block;
785 int s_index, e_index; /* block indexes of the freed allocation */
786 int s_off, e_off; /* block offsets of the freed allocation */
Dennis Zhou (Facebook)b185cd02017-07-24 19:02:17 -0400787 int start, end; /* start and end of the whole free area */
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400788
789 /*
790 * Calculate per block offsets.
791 * The calculation uses an inclusive range, but the resulting offsets
792 * are [start, end). e_index always points to the last block in the
793 * range.
794 */
795 s_index = pcpu_off_to_block_index(bit_off);
796 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
797 s_off = pcpu_off_to_block_off(bit_off);
798 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
799
800 s_block = chunk->md_blocks + s_index;
801 e_block = chunk->md_blocks + e_index;
802
Dennis Zhou (Facebook)b185cd02017-07-24 19:02:17 -0400803 /*
804 * Check if the freed area aligns with the block->contig_hint.
805 * If it does, then the scan to find the beginning/end of the
806 * larger free area can be avoided.
807 *
808 * start and end refer to beginning and end of the free area
809 * within each their respective blocks. This is not necessarily
810 * the entire free area as it may span blocks past the beginning
811 * or end of the block.
812 */
813 start = s_off;
814 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
815 start = s_block->contig_hint_start;
816 } else {
817 /*
818 * Scan backwards to find the extent of the free area.
819 * find_last_bit returns the starting bit, so if the start bit
820 * is returned, that means there was no last bit and the
821 * remainder of the chunk is free.
822 */
823 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
824 start);
825 start = (start == l_bit) ? 0 : l_bit + 1;
826 }
827
828 end = e_off;
829 if (e_off == e_block->contig_hint_start)
830 end = e_block->contig_hint_start + e_block->contig_hint;
831 else
832 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
833 PCPU_BITMAP_BLOCK_BITS, end);
834
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400835 /* update s_block */
Dennis Zhou (Facebook)b185cd02017-07-24 19:02:17 -0400836 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
837 pcpu_block_update(s_block, start, e_off);
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400838
839 /* freeing in the same block */
840 if (s_index != e_index) {
841 /* update e_block */
Dennis Zhou (Facebook)b185cd02017-07-24 19:02:17 -0400842 pcpu_block_update(e_block, 0, end);
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400843
844 /* reset md_blocks in the middle */
845 for (block = s_block + 1; block < e_block; block++) {
846 block->first_free = 0;
847 block->contig_hint_start = 0;
848 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
849 block->left_free = PCPU_BITMAP_BLOCK_BITS;
850 block->right_free = PCPU_BITMAP_BLOCK_BITS;
851 }
852 }
853
Dennis Zhou (Facebook)b185cd02017-07-24 19:02:17 -0400854 /*
855 * Refresh chunk metadata when the free makes a page free, a block
856 * free, or spans across blocks. The contig hint may be off by up to
857 * a page, but if the hint is contained in a block, it will be accurate
858 * with the else condition below.
859 */
860 if ((ALIGN_DOWN(end, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS)) >
861 ALIGN(start, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS))) ||
862 s_index != e_index)
863 pcpu_chunk_refresh_hint(chunk);
864 else
865 pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start),
866 s_block->contig_hint);
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -0400867}
868
869/**
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700870 * pcpu_is_populated - determines if the region is populated
871 * @chunk: chunk of interest
872 * @bit_off: chunk offset
873 * @bits: size of area
874 * @next_off: return value for the next offset to start searching
875 *
876 * For atomic allocations, check if the backing pages are populated.
877 *
878 * RETURNS:
879 * Bool if the backing pages are populated.
880 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
881 */
882static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
883 int *next_off)
884{
885 int page_start, page_end, rs, re;
886
887 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
888 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
889
890 rs = page_start;
891 pcpu_next_unpop(chunk->populated, &rs, &re, page_end);
892 if (rs >= page_end)
893 return true;
894
895 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
896 return false;
897}
898
899/**
900 * pcpu_find_block_fit - finds the block index to start searching
901 * @chunk: chunk of interest
902 * @alloc_bits: size of request in allocation units
903 * @align: alignment of area (max PAGE_SIZE bytes)
904 * @pop_only: use populated regions only
905 *
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400906 * Given a chunk and an allocation spec, find the offset to begin searching
907 * for a free region. This iterates over the bitmap metadata blocks to
908 * find an offset that will be guaranteed to fit the requirements. It is
909 * not quite first fit as if the allocation does not fit in the contig hint
910 * of a block or chunk, it is skipped. This errs on the side of caution
911 * to prevent excess iteration. Poor alignment can cause the allocator to
912 * skip over blocks and chunks that have valid free areas.
913 *
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700914 * RETURNS:
915 * The offset in the bitmap to begin searching.
916 * -1 if no offset is found.
917 */
918static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
919 size_t align, bool pop_only)
920{
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400921 int bit_off, bits, next_off;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700922
Dennis Zhou (Facebook)13f96632017-07-24 19:02:14 -0400923 /*
924 * Check to see if the allocation can fit in the chunk's contig hint.
925 * This is an optimization to prevent scanning by assuming if it
926 * cannot fit in the global hint, there is memory pressure and creating
927 * a new chunk would happen soon.
928 */
929 bit_off = ALIGN(chunk->contig_bits_start, align) -
930 chunk->contig_bits_start;
931 if (bit_off + alloc_bits > chunk->contig_bits)
932 return -1;
933
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400934 bit_off = chunk->first_bit;
935 bits = 0;
936 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700937 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400938 &next_off))
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700939 break;
940
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400941 bit_off = next_off;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700942 bits = 0;
943 }
944
945 if (bit_off == pcpu_chunk_map_bits(chunk))
946 return -1;
947
948 return bit_off;
949}
950
951/**
952 * pcpu_alloc_area - allocates an area from a pcpu_chunk
953 * @chunk: chunk of interest
954 * @alloc_bits: size of request in allocation units
955 * @align: alignment of area (max PAGE_SIZE)
956 * @start: bit_off to start searching
957 *
958 * This function takes in a @start offset to begin searching to fit an
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400959 * allocation of @alloc_bits with alignment @align. It needs to scan
960 * the allocation map because if it fits within the block's contig hint,
961 * @start will be block->first_free. This is an attempt to fill the
962 * allocation prior to breaking the contig hint. The allocation and
963 * boundary maps are updated accordingly if it confirms a valid
964 * free area.
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700965 *
966 * RETURNS:
967 * Allocated addr offset in @chunk on success.
968 * -1 if no matching area is found.
969 */
970static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
971 size_t align, int start)
972{
973 size_t align_mask = (align) ? (align - 1) : 0;
974 int bit_off, end, oslot;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900975
Tejun Heo4f996e22016-05-25 11:48:25 -0400976 lockdep_assert_held(&pcpu_lock);
977
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700978 oslot = pcpu_chunk_slot(chunk);
Tejun Heo833af842009-11-11 15:35:18 +0900979
980 /*
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700981 * Search to find a fit.
Tejun Heo833af842009-11-11 15:35:18 +0900982 */
Dennis Zhou (Facebook)b4c21162017-07-24 19:02:19 -0400983 end = start + alloc_bits + PCPU_BITMAP_BLOCK_BITS;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700984 bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start,
985 alloc_bits, align_mask);
986 if (bit_off >= end)
987 return -1;
Tejun Heo833af842009-11-11 15:35:18 +0900988
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700989 /* update alloc map */
990 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900991
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700992 /* update boundary map */
993 set_bit(bit_off, chunk->bound_map);
994 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
995 set_bit(bit_off + alloc_bits, chunk->bound_map);
Tejun Heoa16037c2014-09-02 14:46:02 -0400996
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -0700997 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
Tejun Heoa16037c2014-09-02 14:46:02 -0400998
Dennis Zhou (Facebook)86b442f2017-07-24 19:02:13 -0400999 /* update first free bit */
1000 if (bit_off == chunk->first_bit)
1001 chunk->first_bit = find_next_zero_bit(
1002 chunk->alloc_map,
1003 pcpu_chunk_map_bits(chunk),
1004 bit_off + alloc_bits);
1005
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001006 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
Tejun Heoa16037c2014-09-02 14:46:02 -04001007
Tejun Heofbf59bc2009-02-20 16:29:08 +09001008 pcpu_chunk_relocate(chunk, oslot);
1009
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001010 return bit_off * PCPU_MIN_ALLOC_SIZE;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001011}
1012
1013/**
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001014 * pcpu_free_area - frees the corresponding offset
Tejun Heofbf59bc2009-02-20 16:29:08 +09001015 * @chunk: chunk of interest
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001016 * @off: addr offset into chunk
Tejun Heofbf59bc2009-02-20 16:29:08 +09001017 *
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001018 * This function determines the size of an allocation to free using
1019 * the boundary bitmap and clears the allocation map.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001020 */
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001021static void pcpu_free_area(struct pcpu_chunk *chunk, int off)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001022{
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001023 int bit_off, bits, end, oslot;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001024
Dennis Zhou5ccd30e2017-06-19 19:28:29 -04001025 lockdep_assert_held(&pcpu_lock);
Dennis Zhou30a5b532017-06-19 19:28:31 -04001026 pcpu_stats_area_dealloc(chunk);
Dennis Zhou5ccd30e2017-06-19 19:28:29 -04001027
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001028 oslot = pcpu_chunk_slot(chunk);
Al Viro723ad1d2014-03-06 21:13:18 -05001029
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001030 bit_off = off / PCPU_MIN_ALLOC_SIZE;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001031
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001032 /* find end index */
1033 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1034 bit_off + 1);
1035 bits = end - bit_off;
1036 bitmap_clear(chunk->alloc_map, bit_off, bits);
Al Viro3d331ad2014-03-06 20:52:32 -05001037
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001038 /* update metadata */
1039 chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001040
Dennis Zhou (Facebook)86b442f2017-07-24 19:02:13 -04001041 /* update first free bit */
1042 chunk->first_bit = min(chunk->first_bit, bit_off);
1043
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001044 pcpu_block_update_hint_free(chunk, bit_off, bits);
Tejun Heob539b872014-09-02 14:46:05 -04001045
Tejun Heofbf59bc2009-02-20 16:29:08 +09001046 pcpu_chunk_relocate(chunk, oslot);
1047}
1048
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001049static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1050{
1051 struct pcpu_block_md *md_block;
1052
1053 for (md_block = chunk->md_blocks;
1054 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1055 md_block++) {
1056 md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1057 md_block->left_free = PCPU_BITMAP_BLOCK_BITS;
1058 md_block->right_free = PCPU_BITMAP_BLOCK_BITS;
1059 }
1060}
1061
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001062/**
1063 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1064 * @tmp_addr: the start of the region served
1065 * @map_size: size of the region served
1066 *
1067 * This is responsible for creating the chunks that serve the first chunk. The
1068 * base_addr is page aligned down of @tmp_addr while the region end is page
1069 * aligned up. Offsets are kept track of to determine the region served. All
1070 * this is done to appease the bitmap allocator in avoiding partial blocks.
1071 *
1072 * RETURNS:
1073 * Chunk serving the region at @tmp_addr of @map_size.
1074 */
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001075static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001076 int map_size)
Dennis Zhou (Facebook)10edf5b2017-07-24 19:02:02 -04001077{
1078 struct pcpu_chunk *chunk;
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001079 unsigned long aligned_addr, lcm_align;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001080 int start_offset, offset_bits, region_size, region_bits;
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001081
1082 /* region calculations */
1083 aligned_addr = tmp_addr & PAGE_MASK;
1084
1085 start_offset = tmp_addr - aligned_addr;
Dennis Zhou (Facebook)6b9d7c82017-07-24 19:02:03 -04001086
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001087 /*
1088 * Align the end of the region with the LCM of PAGE_SIZE and
1089 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1090 * the other.
1091 */
1092 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1093 region_size = ALIGN(start_offset + map_size, lcm_align);
Dennis Zhou (Facebook)10edf5b2017-07-24 19:02:02 -04001094
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001095 /* allocate chunk */
Dennis Zhou (Facebook)8ab16c42017-07-24 19:02:07 -04001096 chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) +
1097 BITS_TO_LONGS(region_size >> PAGE_SHIFT),
1098 0);
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001099
Dennis Zhou (Facebook)10edf5b2017-07-24 19:02:02 -04001100 INIT_LIST_HEAD(&chunk->list);
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001101
1102 chunk->base_addr = (void *)aligned_addr;
Dennis Zhou (Facebook)10edf5b2017-07-24 19:02:02 -04001103 chunk->start_offset = start_offset;
Dennis Zhou (Facebook)6b9d7c82017-07-24 19:02:03 -04001104 chunk->end_offset = region_size - chunk->start_offset - map_size;
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001105
Dennis Zhou (Facebook)8ab16c42017-07-24 19:02:07 -04001106 chunk->nr_pages = region_size >> PAGE_SHIFT;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001107 region_bits = pcpu_chunk_map_bits(chunk);
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001108
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001109 chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) *
1110 sizeof(chunk->alloc_map[0]), 0);
1111 chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) *
1112 sizeof(chunk->bound_map[0]), 0);
1113 chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) *
1114 sizeof(chunk->md_blocks[0]), 0);
1115 pcpu_init_md_blocks(chunk);
Dennis Zhou (Facebook)10edf5b2017-07-24 19:02:02 -04001116
1117 /* manage populated page bitmap */
1118 chunk->immutable = true;
Dennis Zhou (Facebook)8ab16c42017-07-24 19:02:07 -04001119 bitmap_fill(chunk->populated, chunk->nr_pages);
1120 chunk->nr_populated = chunk->nr_pages;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001121 chunk->nr_empty_pop_pages =
1122 pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE,
1123 map_size / PCPU_MIN_ALLOC_SIZE);
Dennis Zhou (Facebook)10edf5b2017-07-24 19:02:02 -04001124
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001125 chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE;
1126 chunk->free_bytes = map_size;
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001127
1128 if (chunk->start_offset) {
1129 /* hide the beginning of the bitmap */
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001130 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1131 bitmap_set(chunk->alloc_map, 0, offset_bits);
1132 set_bit(0, chunk->bound_map);
1133 set_bit(offset_bits, chunk->bound_map);
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001134
Dennis Zhou (Facebook)86b442f2017-07-24 19:02:13 -04001135 chunk->first_bit = offset_bits;
1136
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001137 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001138 }
1139
Dennis Zhou (Facebook)6b9d7c82017-07-24 19:02:03 -04001140 if (chunk->end_offset) {
1141 /* hide the end of the bitmap */
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001142 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1143 bitmap_set(chunk->alloc_map,
1144 pcpu_chunk_map_bits(chunk) - offset_bits,
1145 offset_bits);
1146 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1147 chunk->bound_map);
1148 set_bit(region_bits, chunk->bound_map);
Dennis Zhou (Facebook)6b9d7c82017-07-24 19:02:03 -04001149
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001150 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1151 - offset_bits, offset_bits);
1152 }
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001153
Dennis Zhou (Facebook)10edf5b2017-07-24 19:02:02 -04001154 return chunk;
1155}
1156
Tejun Heo60810892010-04-09 18:57:01 +09001157static struct pcpu_chunk *pcpu_alloc_chunk(void)
1158{
1159 struct pcpu_chunk *chunk;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001160 int region_bits;
Tejun Heo60810892010-04-09 18:57:01 +09001161
Bob Liu90459ce02011-08-04 11:02:33 +02001162 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +09001163 if (!chunk)
1164 return NULL;
1165
Tejun Heo60810892010-04-09 18:57:01 +09001166 INIT_LIST_HEAD(&chunk->list);
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001167 chunk->nr_pages = pcpu_unit_pages;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001168 region_bits = pcpu_chunk_map_bits(chunk);
1169
1170 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1171 sizeof(chunk->alloc_map[0]));
1172 if (!chunk->alloc_map)
1173 goto alloc_map_fail;
1174
1175 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1176 sizeof(chunk->bound_map[0]));
1177 if (!chunk->bound_map)
1178 goto bound_map_fail;
1179
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001180 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1181 sizeof(chunk->md_blocks[0]));
1182 if (!chunk->md_blocks)
1183 goto md_blocks_fail;
1184
1185 pcpu_init_md_blocks(chunk);
1186
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001187 /* init metadata */
1188 chunk->contig_bits = region_bits;
1189 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001190
Tejun Heo60810892010-04-09 18:57:01 +09001191 return chunk;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001192
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04001193md_blocks_fail:
1194 pcpu_mem_free(chunk->bound_map);
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001195bound_map_fail:
1196 pcpu_mem_free(chunk->alloc_map);
1197alloc_map_fail:
1198 pcpu_mem_free(chunk);
1199
1200 return NULL;
Tejun Heo60810892010-04-09 18:57:01 +09001201}
1202
1203static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1204{
1205 if (!chunk)
1206 return;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001207 pcpu_mem_free(chunk->bound_map);
1208 pcpu_mem_free(chunk->alloc_map);
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -08001209 pcpu_mem_free(chunk);
Tejun Heo60810892010-04-09 18:57:01 +09001210}
1211
Tejun Heob539b872014-09-02 14:46:05 -04001212/**
1213 * pcpu_chunk_populated - post-population bookkeeping
1214 * @chunk: pcpu_chunk which got populated
1215 * @page_start: the start page
1216 * @page_end: the end page
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001217 * @for_alloc: if this is to populate for allocation
Tejun Heob539b872014-09-02 14:46:05 -04001218 *
1219 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1220 * the bookkeeping information accordingly. Must be called after each
1221 * successful population.
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001222 *
1223 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1224 * is to serve an allocation in that area.
Tejun Heob539b872014-09-02 14:46:05 -04001225 */
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001226static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1227 int page_end, bool for_alloc)
Tejun Heob539b872014-09-02 14:46:05 -04001228{
1229 int nr = page_end - page_start;
1230
1231 lockdep_assert_held(&pcpu_lock);
1232
1233 bitmap_set(chunk->populated, page_start, nr);
1234 chunk->nr_populated += nr;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001235
1236 if (!for_alloc) {
1237 chunk->nr_empty_pop_pages += nr;
1238 pcpu_nr_empty_pop_pages += nr;
1239 }
Tejun Heob539b872014-09-02 14:46:05 -04001240}
1241
1242/**
1243 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1244 * @chunk: pcpu_chunk which got depopulated
1245 * @page_start: the start page
1246 * @page_end: the end page
1247 *
1248 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1249 * Update the bookkeeping information accordingly. Must be called after
1250 * each successful depopulation.
1251 */
1252static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1253 int page_start, int page_end)
1254{
1255 int nr = page_end - page_start;
1256
1257 lockdep_assert_held(&pcpu_lock);
1258
1259 bitmap_clear(chunk->populated, page_start, nr);
1260 chunk->nr_populated -= nr;
Dennis Zhou (Facebook)0cecf502017-07-24 19:02:08 -04001261 chunk->nr_empty_pop_pages -= nr;
Tejun Heob539b872014-09-02 14:46:05 -04001262 pcpu_nr_empty_pop_pages -= nr;
1263}
1264
Tejun Heo9f645532010-04-09 18:57:01 +09001265/*
1266 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001267 *
Tejun Heo9f645532010-04-09 18:57:01 +09001268 * To allow different implementations, chunk alloc/free and
1269 * [de]population are implemented in a separate file which is pulled
1270 * into this file and compiled together. The following functions
1271 * should be implemented.
Tejun Heoce3141a2009-07-04 08:11:00 +09001272 *
Tejun Heo9f645532010-04-09 18:57:01 +09001273 * pcpu_populate_chunk - populate the specified range of a chunk
1274 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1275 * pcpu_create_chunk - create a new chunk
1276 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1277 * pcpu_addr_to_page - translate address to physical address
1278 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
Tejun Heofbf59bc2009-02-20 16:29:08 +09001279 */
Tejun Heo9f645532010-04-09 18:57:01 +09001280static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
1281static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
1282static struct pcpu_chunk *pcpu_create_chunk(void);
1283static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1284static struct page *pcpu_addr_to_page(void *addr);
1285static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +09001286
Tejun Heob0c97782010-04-09 18:57:01 +09001287#ifdef CONFIG_NEED_PER_CPU_KM
1288#include "percpu-km.c"
1289#else
Tejun Heo9f645532010-04-09 18:57:01 +09001290#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +09001291#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +09001292
1293/**
Tejun Heo88999a82010-04-09 18:57:01 +09001294 * pcpu_chunk_addr_search - determine chunk containing specified address
1295 * @addr: address for which the chunk needs to be determined.
1296 *
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001297 * This is an internal function that handles all but static allocations.
1298 * Static percpu address values should never be passed into the allocator.
1299 *
Tejun Heo88999a82010-04-09 18:57:01 +09001300 * RETURNS:
1301 * The address of the found chunk.
1302 */
1303static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1304{
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001305 /* is it in the dynamic region (first chunk)? */
Dennis Zhou (Facebook)560f2c22017-07-24 19:02:06 -04001306 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
Tejun Heo88999a82010-04-09 18:57:01 +09001307 return pcpu_first_chunk;
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001308
1309 /* is it in the reserved region? */
Dennis Zhou (Facebook)560f2c22017-07-24 19:02:06 -04001310 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001311 return pcpu_reserved_chunk;
Tejun Heo88999a82010-04-09 18:57:01 +09001312
1313 /*
1314 * The address is relative to unit0 which might be unused and
1315 * thus unmapped. Offset the address to the unit space of the
1316 * current processor before looking it up in the vmalloc
1317 * space. Note that any possible cpu id can be used here, so
1318 * there's no need to worry about preemption or cpu hotplug.
1319 */
1320 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +09001321 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +09001322}
1323
1324/**
Tejun Heoedcb4632009-03-06 14:33:59 +09001325 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +09001326 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +09001327 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +09001328 * @reserved: allocate from the reserved chunk if available
Tejun Heo5835d962014-09-02 14:46:04 -04001329 * @gfp: allocation flags
Tejun Heofbf59bc2009-02-20 16:29:08 +09001330 *
Tejun Heo5835d962014-09-02 14:46:04 -04001331 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
Daniel Borkmann0ea7eee2017-10-17 16:55:52 +02001332 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1333 * then no warning will be triggered on invalid or failed allocation
1334 * requests.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001335 *
1336 * RETURNS:
1337 * Percpu pointer to the allocated area on success, NULL on failure.
1338 */
Tejun Heo5835d962014-09-02 14:46:04 -04001339static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1340 gfp_t gfp)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001341{
Daniel Borkmann0ea7eee2017-10-17 16:55:52 +02001342 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1343 bool do_warn = !(gfp & __GFP_NOWARN);
Tejun Heof2badb02009-09-29 09:17:58 +09001344 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001345 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +09001346 const char *err;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001347 int slot, off, cpu, ret;
Jiri Kosina403a91b2009-10-29 00:25:59 +09001348 unsigned long flags;
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001349 void __percpu *ptr;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001350 size_t bits, bit_align;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001351
Al Viro723ad1d2014-03-06 21:13:18 -05001352 /*
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001353 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1354 * therefore alignment must be a minimum of that many bytes.
1355 * An allocation may have internal fragmentation from rounding up
1356 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
Al Viro723ad1d2014-03-06 21:13:18 -05001357 */
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04001358 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1359 align = PCPU_MIN_ALLOC_SIZE;
Al Viro723ad1d2014-03-06 21:13:18 -05001360
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04001361 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001362 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1363 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
Viro2f69fa82014-03-17 16:01:27 -04001364
zijun_hu3ca45a42016-10-14 15:12:54 +08001365 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1366 !is_power_of_2(align))) {
Daniel Borkmann0ea7eee2017-10-17 16:55:52 +02001367 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
Joe Perches756a0252016-03-17 14:19:47 -07001368 size, align);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001369 return NULL;
1370 }
1371
Tejun Heo6710e592016-05-25 11:48:25 -04001372 if (!is_atomic)
1373 mutex_lock(&pcpu_alloc_mutex);
1374
Jiri Kosina403a91b2009-10-29 00:25:59 +09001375 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001376
Tejun Heoedcb4632009-03-06 14:33:59 +09001377 /* serve reserved allocations from the reserved chunk if available */
1378 if (reserved && pcpu_reserved_chunk) {
1379 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +09001380
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001381 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1382 if (off < 0) {
Tejun Heo833af842009-11-11 15:35:18 +09001383 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +09001384 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +09001385 }
Tejun Heo833af842009-11-11 15:35:18 +09001386
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001387 off = pcpu_alloc_area(chunk, bits, bit_align, off);
Tejun Heoedcb4632009-03-06 14:33:59 +09001388 if (off >= 0)
1389 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +09001390
Tejun Heof2badb02009-09-29 09:17:58 +09001391 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +09001392 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +09001393 }
1394
Tejun Heoccea34b2009-03-07 00:44:13 +09001395restart:
Tejun Heoedcb4632009-03-06 14:33:59 +09001396 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001397 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1398 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001399 off = pcpu_find_block_fit(chunk, bits, bit_align,
1400 is_atomic);
1401 if (off < 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001402 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +09001403
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001404 off = pcpu_alloc_area(chunk, bits, bit_align, off);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001405 if (off >= 0)
1406 goto area_found;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001407
Tejun Heofbf59bc2009-02-20 16:29:08 +09001408 }
1409 }
1410
Jiri Kosina403a91b2009-10-29 00:25:59 +09001411 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001412
Tejun Heob38d08f2014-09-02 14:46:02 -04001413 /*
1414 * No space left. Create a new chunk. We don't want multiple
1415 * tasks to create chunks simultaneously. Serialize and create iff
1416 * there's still no empty chunk after grabbing the mutex.
1417 */
Dennis Zhou11df02b2017-06-21 11:51:09 -04001418 if (is_atomic) {
1419 err = "atomic alloc failed, no space left";
Tejun Heo5835d962014-09-02 14:46:04 -04001420 goto fail;
Dennis Zhou11df02b2017-06-21 11:51:09 -04001421 }
Tejun Heo5835d962014-09-02 14:46:04 -04001422
Tejun Heob38d08f2014-09-02 14:46:02 -04001423 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1424 chunk = pcpu_create_chunk();
1425 if (!chunk) {
1426 err = "failed to allocate new chunk";
1427 goto fail;
1428 }
1429
1430 spin_lock_irqsave(&pcpu_lock, flags);
1431 pcpu_chunk_relocate(chunk, -1);
1432 } else {
1433 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heof2badb02009-09-29 09:17:58 +09001434 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001435
Tejun Heoccea34b2009-03-07 00:44:13 +09001436 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001437
1438area_found:
Dennis Zhou30a5b532017-06-19 19:28:31 -04001439 pcpu_stats_area_alloc(chunk, size);
Jiri Kosina403a91b2009-10-29 00:25:59 +09001440 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001441
Tejun Heodca49642014-09-02 14:46:01 -04001442 /* populate if not all pages are already there */
Tejun Heo5835d962014-09-02 14:46:04 -04001443 if (!is_atomic) {
Tejun Heoe04d3202014-09-02 14:46:04 -04001444 int page_start, page_end, rs, re;
Tejun Heodca49642014-09-02 14:46:01 -04001445
Tejun Heoe04d3202014-09-02 14:46:04 -04001446 page_start = PFN_DOWN(off);
1447 page_end = PFN_UP(off + size);
Tejun Heob38d08f2014-09-02 14:46:02 -04001448
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -04001449 pcpu_for_each_unpop_region(chunk->populated, rs, re,
1450 page_start, page_end) {
Tejun Heoe04d3202014-09-02 14:46:04 -04001451 WARN_ON(chunk->immutable);
1452
1453 ret = pcpu_populate_chunk(chunk, rs, re);
1454
1455 spin_lock_irqsave(&pcpu_lock, flags);
1456 if (ret) {
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001457 pcpu_free_area(chunk, off);
Tejun Heoe04d3202014-09-02 14:46:04 -04001458 err = "failed to populate";
1459 goto fail_unlock;
1460 }
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001461 pcpu_chunk_populated(chunk, rs, re, true);
Tejun Heoe04d3202014-09-02 14:46:04 -04001462 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heodca49642014-09-02 14:46:01 -04001463 }
Tejun Heofbf59bc2009-02-20 16:29:08 +09001464
Tejun Heoe04d3202014-09-02 14:46:04 -04001465 mutex_unlock(&pcpu_alloc_mutex);
1466 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001467
Tejun Heo1a4d7602014-09-02 14:46:05 -04001468 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1469 pcpu_schedule_balance_work();
1470
Tejun Heodca49642014-09-02 14:46:01 -04001471 /* clear the areas and return address relative to base address */
1472 for_each_possible_cpu(cpu)
1473 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1474
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001475 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
Larry Finger8a8c35f2015-06-24 16:58:51 -07001476 kmemleak_alloc_percpu(ptr, size, gfp);
Dennis Zhoudf95e792017-06-19 19:28:32 -04001477
1478 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1479 chunk->base_addr, off, ptr);
1480
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001481 return ptr;
Tejun Heoccea34b2009-03-07 00:44:13 +09001482
1483fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001484 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heob38d08f2014-09-02 14:46:02 -04001485fail:
Dennis Zhoudf95e792017-06-19 19:28:32 -04001486 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1487
Daniel Borkmann0ea7eee2017-10-17 16:55:52 +02001488 if (!is_atomic && do_warn && warn_limit) {
Joe Perches870d4b12016-03-17 14:19:53 -07001489 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
Joe Perches598d8092016-03-17 14:19:44 -07001490 size, align, is_atomic, err);
Tejun Heof2badb02009-09-29 09:17:58 +09001491 dump_stack();
1492 if (!--warn_limit)
Joe Perches870d4b12016-03-17 14:19:53 -07001493 pr_info("limit reached, disable warning\n");
Tejun Heof2badb02009-09-29 09:17:58 +09001494 }
Tejun Heo1a4d7602014-09-02 14:46:05 -04001495 if (is_atomic) {
1496 /* see the flag handling in pcpu_blance_workfn() */
1497 pcpu_atomic_alloc_failed = true;
1498 pcpu_schedule_balance_work();
Tejun Heo6710e592016-05-25 11:48:25 -04001499 } else {
1500 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heo1a4d7602014-09-02 14:46:05 -04001501 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001502 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001503}
Tejun Heoedcb4632009-03-06 14:33:59 +09001504
1505/**
Tejun Heo5835d962014-09-02 14:46:04 -04001506 * __alloc_percpu_gfp - allocate dynamic percpu area
Tejun Heoedcb4632009-03-06 14:33:59 +09001507 * @size: size of area to allocate in bytes
1508 * @align: alignment of area (max PAGE_SIZE)
Tejun Heo5835d962014-09-02 14:46:04 -04001509 * @gfp: allocation flags
Tejun Heoedcb4632009-03-06 14:33:59 +09001510 *
Tejun Heo5835d962014-09-02 14:46:04 -04001511 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1512 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
Daniel Borkmann0ea7eee2017-10-17 16:55:52 +02001513 * be called from any context but is a lot more likely to fail. If @gfp
1514 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1515 * allocation requests.
Tejun Heoccea34b2009-03-07 00:44:13 +09001516 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001517 * RETURNS:
1518 * Percpu pointer to the allocated area on success, NULL on failure.
1519 */
Tejun Heo5835d962014-09-02 14:46:04 -04001520void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1521{
1522 return pcpu_alloc(size, align, false, gfp);
1523}
1524EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1525
1526/**
1527 * __alloc_percpu - allocate dynamic percpu area
1528 * @size: size of area to allocate in bytes
1529 * @align: alignment of area (max PAGE_SIZE)
1530 *
1531 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1532 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001533void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +09001534{
Tejun Heo5835d962014-09-02 14:46:04 -04001535 return pcpu_alloc(size, align, false, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +09001536}
Tejun Heofbf59bc2009-02-20 16:29:08 +09001537EXPORT_SYMBOL_GPL(__alloc_percpu);
1538
Tejun Heoedcb4632009-03-06 14:33:59 +09001539/**
1540 * __alloc_reserved_percpu - allocate reserved percpu area
1541 * @size: size of area to allocate in bytes
1542 * @align: alignment of area (max PAGE_SIZE)
1543 *
Tejun Heo9329ba92010-09-10 11:01:56 +02001544 * Allocate zero-filled percpu area of @size bytes aligned at @align
1545 * from reserved percpu area if arch has set it up; otherwise,
1546 * allocation is served from the same dynamic area. Might sleep.
1547 * Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +09001548 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001549 * CONTEXT:
1550 * Does GFP_KERNEL allocation.
1551 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001552 * RETURNS:
1553 * Percpu pointer to the allocated area on success, NULL on failure.
1554 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001555void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +09001556{
Tejun Heo5835d962014-09-02 14:46:04 -04001557 return pcpu_alloc(size, align, true, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +09001558}
1559
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001560/**
Tejun Heo1a4d7602014-09-02 14:46:05 -04001561 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001562 * @work: unused
1563 *
1564 * Reclaim all fully free chunks except for the first one.
1565 */
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001566static void pcpu_balance_workfn(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001567{
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001568 LIST_HEAD(to_free);
1569 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001570 struct pcpu_chunk *chunk, *next;
Tejun Heo1a4d7602014-09-02 14:46:05 -04001571 int slot, nr_to_pop, ret;
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001572
Tejun Heo1a4d7602014-09-02 14:46:05 -04001573 /*
1574 * There's no reason to keep around multiple unused chunks and VM
1575 * areas can be scarce. Destroy all free chunks except for one.
1576 */
Tejun Heoccea34b2009-03-07 00:44:13 +09001577 mutex_lock(&pcpu_alloc_mutex);
1578 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001579
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001580 list_for_each_entry_safe(chunk, next, free_head, list) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001581 WARN_ON(chunk->immutable);
1582
1583 /* spare the first one */
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001584 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001585 continue;
1586
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001587 list_move(&chunk->list, &to_free);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001588 }
1589
Tejun Heoccea34b2009-03-07 00:44:13 +09001590 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001591
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001592 list_for_each_entry_safe(chunk, next, &to_free, list) {
Tejun Heoa93ace42014-09-02 14:46:02 -04001593 int rs, re;
Tejun Heodca49642014-09-02 14:46:01 -04001594
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -04001595 pcpu_for_each_pop_region(chunk->populated, rs, re, 0,
1596 chunk->nr_pages) {
Tejun Heoa93ace42014-09-02 14:46:02 -04001597 pcpu_depopulate_chunk(chunk, rs, re);
Tejun Heob539b872014-09-02 14:46:05 -04001598 spin_lock_irq(&pcpu_lock);
1599 pcpu_chunk_depopulated(chunk, rs, re);
1600 spin_unlock_irq(&pcpu_lock);
Tejun Heoa93ace42014-09-02 14:46:02 -04001601 }
Tejun Heo60810892010-04-09 18:57:01 +09001602 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001603 }
Tejun Heo971f3912009-08-14 15:00:49 +09001604
Tejun Heo1a4d7602014-09-02 14:46:05 -04001605 /*
1606 * Ensure there are certain number of free populated pages for
1607 * atomic allocs. Fill up from the most packed so that atomic
1608 * allocs don't increase fragmentation. If atomic allocation
1609 * failed previously, always populate the maximum amount. This
1610 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1611 * failing indefinitely; however, large atomic allocs are not
1612 * something we support properly and can be highly unreliable and
1613 * inefficient.
1614 */
1615retry_pop:
1616 if (pcpu_atomic_alloc_failed) {
1617 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1618 /* best effort anyway, don't worry about synchronization */
1619 pcpu_atomic_alloc_failed = false;
1620 } else {
1621 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1622 pcpu_nr_empty_pop_pages,
1623 0, PCPU_EMPTY_POP_PAGES_HIGH);
1624 }
1625
1626 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1627 int nr_unpop = 0, rs, re;
1628
1629 if (!nr_to_pop)
1630 break;
1631
1632 spin_lock_irq(&pcpu_lock);
1633 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
Dennis Zhou (Facebook)8ab16c42017-07-24 19:02:07 -04001634 nr_unpop = chunk->nr_pages - chunk->nr_populated;
Tejun Heo1a4d7602014-09-02 14:46:05 -04001635 if (nr_unpop)
1636 break;
1637 }
1638 spin_unlock_irq(&pcpu_lock);
1639
1640 if (!nr_unpop)
1641 continue;
1642
1643 /* @chunk can't go away while pcpu_alloc_mutex is held */
Dennis Zhou (Facebook)91e914c2017-07-24 19:02:10 -04001644 pcpu_for_each_unpop_region(chunk->populated, rs, re, 0,
1645 chunk->nr_pages) {
Tejun Heo1a4d7602014-09-02 14:46:05 -04001646 int nr = min(re - rs, nr_to_pop);
1647
1648 ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1649 if (!ret) {
1650 nr_to_pop -= nr;
1651 spin_lock_irq(&pcpu_lock);
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001652 pcpu_chunk_populated(chunk, rs, rs + nr, false);
Tejun Heo1a4d7602014-09-02 14:46:05 -04001653 spin_unlock_irq(&pcpu_lock);
1654 } else {
1655 nr_to_pop = 0;
1656 }
1657
1658 if (!nr_to_pop)
1659 break;
1660 }
1661 }
1662
1663 if (nr_to_pop) {
1664 /* ran out of chunks to populate, create a new one and retry */
1665 chunk = pcpu_create_chunk();
1666 if (chunk) {
1667 spin_lock_irq(&pcpu_lock);
1668 pcpu_chunk_relocate(chunk, -1);
1669 spin_unlock_irq(&pcpu_lock);
1670 goto retry_pop;
1671 }
1672 }
1673
Tejun Heo971f3912009-08-14 15:00:49 +09001674 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001675}
1676
1677/**
1678 * free_percpu - free percpu area
1679 * @ptr: pointer to area to free
1680 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001681 * Free percpu area @ptr.
1682 *
1683 * CONTEXT:
1684 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001685 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001686void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001687{
Andrew Morton129182e2010-01-08 14:42:39 -08001688 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001689 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +09001690 unsigned long flags;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001691 int off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001692
1693 if (!ptr)
1694 return;
1695
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001696 kmemleak_free_percpu(ptr);
1697
Andrew Morton129182e2010-01-08 14:42:39 -08001698 addr = __pcpu_ptr_to_addr(ptr);
1699
Tejun Heoccea34b2009-03-07 00:44:13 +09001700 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001701
1702 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +09001703 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001704
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001705 pcpu_free_area(chunk, off);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001706
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001707 /* if there are more than one fully free chunks, wake up grim reaper */
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07001708 if (chunk->free_bytes == pcpu_unit_size) {
Tejun Heofbf59bc2009-02-20 16:29:08 +09001709 struct pcpu_chunk *pos;
1710
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001711 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001712 if (pos != chunk) {
Tejun Heo1a4d7602014-09-02 14:46:05 -04001713 pcpu_schedule_balance_work();
Tejun Heofbf59bc2009-02-20 16:29:08 +09001714 break;
1715 }
1716 }
1717
Dennis Zhoudf95e792017-06-19 19:28:32 -04001718 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1719
Tejun Heoccea34b2009-03-07 00:44:13 +09001720 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001721}
1722EXPORT_SYMBOL_GPL(free_percpu);
1723
Thomas Gleixner383776f2017-02-27 15:37:36 +01001724bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1725{
1726#ifdef CONFIG_SMP
1727 const size_t static_size = __per_cpu_end - __per_cpu_start;
1728 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1729 unsigned int cpu;
1730
1731 for_each_possible_cpu(cpu) {
1732 void *start = per_cpu_ptr(base, cpu);
1733 void *va = (void *)addr;
1734
1735 if (va >= start && va < start + static_size) {
Peter Zijlstra8ce371f2017-03-20 12:26:55 +01001736 if (can_addr) {
Thomas Gleixner383776f2017-02-27 15:37:36 +01001737 *can_addr = (unsigned long) (va - start);
Peter Zijlstra8ce371f2017-03-20 12:26:55 +01001738 *can_addr += (unsigned long)
1739 per_cpu_ptr(base, get_boot_cpu_id());
1740 }
Thomas Gleixner383776f2017-02-27 15:37:36 +01001741 return true;
1742 }
1743 }
1744#endif
1745 /* on UP, can't distinguish from other static vars, always false */
1746 return false;
1747}
1748
Vivek Goyal3b034b02009-11-24 15:50:03 +09001749/**
Tejun Heo10fad5e2010-03-10 18:57:54 +09001750 * is_kernel_percpu_address - test whether address is from static percpu area
1751 * @addr: address to test
1752 *
1753 * Test whether @addr belongs to in-kernel static percpu area. Module
1754 * static percpu areas are not considered. For those, use
1755 * is_module_percpu_address().
1756 *
1757 * RETURNS:
1758 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1759 */
1760bool is_kernel_percpu_address(unsigned long addr)
1761{
Thomas Gleixner383776f2017-02-27 15:37:36 +01001762 return __is_kernel_percpu_address(addr, NULL);
Tejun Heo10fad5e2010-03-10 18:57:54 +09001763}
1764
1765/**
Vivek Goyal3b034b02009-11-24 15:50:03 +09001766 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1767 * @addr: the address to be converted to physical address
1768 *
1769 * Given @addr which is dereferenceable address obtained via one of
1770 * percpu access macros, this function translates it into its physical
1771 * address. The caller is responsible for ensuring @addr stays valid
1772 * until this function finishes.
1773 *
Dave Young67589c72011-11-23 08:20:53 -08001774 * percpu allocator has special setup for the first chunk, which currently
1775 * supports either embedding in linear address space or vmalloc mapping,
1776 * and, from the second one, the backing allocator (currently either vm or
1777 * km) provides translation.
1778 *
Yannick Guerrinibffc4372015-03-06 23:30:42 +01001779 * The addr can be translated simply without checking if it falls into the
Dave Young67589c72011-11-23 08:20:53 -08001780 * first chunk. But the current code reflects better how percpu allocator
1781 * actually works, and the verification can discover both bugs in percpu
1782 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1783 * code.
1784 *
Vivek Goyal3b034b02009-11-24 15:50:03 +09001785 * RETURNS:
1786 * The physical address for @addr.
1787 */
1788phys_addr_t per_cpu_ptr_to_phys(void *addr)
1789{
Tejun Heo9983b6f02010-06-18 11:44:31 +02001790 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1791 bool in_first_chunk = false;
Tejun Heoa855b842011-11-18 10:55:35 -08001792 unsigned long first_low, first_high;
Tejun Heo9983b6f02010-06-18 11:44:31 +02001793 unsigned int cpu;
1794
1795 /*
Tejun Heoa855b842011-11-18 10:55:35 -08001796 * The following test on unit_low/high isn't strictly
Tejun Heo9983b6f02010-06-18 11:44:31 +02001797 * necessary but will speed up lookups of addresses which
1798 * aren't in the first chunk.
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001799 *
1800 * The address check is against full chunk sizes. pcpu_base_addr
1801 * points to the beginning of the first chunk including the
1802 * static region. Assumes good intent as the first chunk may
1803 * not be full (ie. < pcpu_unit_pages in size).
Tejun Heo9983b6f02010-06-18 11:44:31 +02001804 */
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001805 first_low = (unsigned long)pcpu_base_addr +
1806 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
1807 first_high = (unsigned long)pcpu_base_addr +
1808 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
Tejun Heoa855b842011-11-18 10:55:35 -08001809 if ((unsigned long)addr >= first_low &&
1810 (unsigned long)addr < first_high) {
Tejun Heo9983b6f02010-06-18 11:44:31 +02001811 for_each_possible_cpu(cpu) {
1812 void *start = per_cpu_ptr(base, cpu);
1813
1814 if (addr >= start && addr < start + pcpu_unit_size) {
1815 in_first_chunk = true;
1816 break;
1817 }
1818 }
1819 }
1820
1821 if (in_first_chunk) {
David Howellseac522e2011-03-28 12:53:29 +01001822 if (!is_vmalloc_addr(addr))
Tejun Heo020ec652010-04-09 18:57:00 +09001823 return __pa(addr);
1824 else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001825 return page_to_phys(vmalloc_to_page(addr)) +
1826 offset_in_page(addr);
Tejun Heo020ec652010-04-09 18:57:00 +09001827 } else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001828 return page_to_phys(pcpu_addr_to_page(addr)) +
1829 offset_in_page(addr);
Vivek Goyal3b034b02009-11-24 15:50:03 +09001830}
1831
Tejun Heofbf59bc2009-02-20 16:29:08 +09001832/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001833 * pcpu_alloc_alloc_info - allocate percpu allocation info
1834 * @nr_groups: the number of groups
1835 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001836 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001837 * Allocate ai which is large enough for @nr_groups groups containing
1838 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1839 * cpu_map array which is long enough for @nr_units and filled with
1840 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1841 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001842 *
1843 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001844 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1845 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001846 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001847struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1848 int nr_units)
1849{
1850 struct pcpu_alloc_info *ai;
1851 size_t base_size, ai_size;
1852 void *ptr;
1853 int unit;
1854
1855 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1856 __alignof__(ai->groups[0].cpu_map[0]));
1857 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1858
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001859 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001860 if (!ptr)
1861 return NULL;
1862 ai = ptr;
1863 ptr += base_size;
1864
1865 ai->groups[0].cpu_map = ptr;
1866
1867 for (unit = 0; unit < nr_units; unit++)
1868 ai->groups[0].cpu_map[unit] = NR_CPUS;
1869
1870 ai->nr_groups = nr_groups;
1871 ai->__ai_size = PFN_ALIGN(ai_size);
1872
1873 return ai;
1874}
1875
1876/**
1877 * pcpu_free_alloc_info - free percpu allocation info
1878 * @ai: pcpu_alloc_info to free
1879 *
1880 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1881 */
1882void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1883{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001884 memblock_free_early(__pa(ai), ai->__ai_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001885}
1886
1887/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001888 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1889 * @lvl: loglevel
1890 * @ai: allocation info to dump
1891 *
1892 * Print out information about @ai using loglevel @lvl.
1893 */
1894static void pcpu_dump_alloc_info(const char *lvl,
1895 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001896{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001897 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001898 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001899 int alloc = 0, alloc_end = 0;
1900 int group, v;
1901 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001902
Tejun Heofd1e8a12009-08-14 15:00:51 +09001903 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001904 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001905 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001906
Tejun Heofd1e8a12009-08-14 15:00:51 +09001907 v = num_possible_cpus();
1908 while (v /= 10)
1909 cpu_width++;
1910 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001911
Tejun Heofd1e8a12009-08-14 15:00:51 +09001912 upa = ai->alloc_size / ai->unit_size;
1913 width = upa * (cpu_width + 1) + group_width + 3;
1914 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001915
Tejun Heofd1e8a12009-08-14 15:00:51 +09001916 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1917 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1918 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1919
1920 for (group = 0; group < ai->nr_groups; group++) {
1921 const struct pcpu_group_info *gi = &ai->groups[group];
1922 int unit = 0, unit_end = 0;
1923
1924 BUG_ON(gi->nr_units % upa);
1925 for (alloc_end += gi->nr_units / upa;
1926 alloc < alloc_end; alloc++) {
1927 if (!(alloc % apl)) {
Joe Perches11705322016-03-17 14:19:50 -07001928 pr_cont("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001929 printk("%spcpu-alloc: ", lvl);
1930 }
Joe Perches11705322016-03-17 14:19:50 -07001931 pr_cont("[%0*d] ", group_width, group);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001932
1933 for (unit_end += upa; unit < unit_end; unit++)
1934 if (gi->cpu_map[unit] != NR_CPUS)
Joe Perches11705322016-03-17 14:19:50 -07001935 pr_cont("%0*d ",
1936 cpu_width, gi->cpu_map[unit]);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001937 else
Joe Perches11705322016-03-17 14:19:50 -07001938 pr_cont("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001939 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001940 }
Joe Perches11705322016-03-17 14:19:50 -07001941 pr_cont("\n");
Tejun Heo033e48f2009-08-14 15:00:51 +09001942}
Tejun Heo033e48f2009-08-14 15:00:51 +09001943
Tejun Heofbf59bc2009-02-20 16:29:08 +09001944/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001945 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001946 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001947 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001948 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001949 * Initialize the first percpu chunk which contains the kernel static
1950 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001951 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001952 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001953 * @ai contains all information necessary to initialize the first
1954 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001955 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001956 * @ai->static_size is the size of static percpu area.
1957 *
1958 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001959 * reserve after the static area in the first chunk. This reserves
1960 * the first chunk such that it's available only through reserved
1961 * percpu allocation. This is primarily used to serve module percpu
1962 * static areas on architectures where the addressing model has
1963 * limited offset range for symbol relocations to guarantee module
1964 * percpu symbols fall inside the relocatable range.
1965 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001966 * @ai->dyn_size determines the number of bytes available for dynamic
1967 * allocation in the first chunk. The area between @ai->static_size +
1968 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001969 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001970 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1971 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1972 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001973 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001974 * @ai->atom_size is the allocation atom size and used as alignment
1975 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001976 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001977 * @ai->alloc_size is the allocation size and always multiple of
1978 * @ai->atom_size. This is larger than @ai->atom_size if
1979 * @ai->unit_size is larger than @ai->atom_size.
1980 *
1981 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1982 * percpu areas. Units which should be colocated are put into the
1983 * same group. Dynamic VM areas will be allocated according to these
1984 * groupings. If @ai->nr_groups is zero, a single group containing
1985 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001986 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001987 * The caller should have mapped the first chunk at @base_addr and
1988 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001989 *
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04001990 * The first chunk will always contain a static and a dynamic region.
1991 * However, the static region is not managed by any chunk. If the first
1992 * chunk also contains a reserved region, it is served by two chunks -
1993 * one for the reserved region and one for the dynamic region. They
1994 * share the same vm, but use offset regions in the area allocation map.
1995 * The chunk serving the dynamic region is circulated in the chunk slots
1996 * and available for dynamic allocation like any other chunk.
Tejun Heoedcb4632009-03-06 14:33:59 +09001997 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001998 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001999 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09002000 */
Tejun Heofb435d52009-08-14 15:00:51 +09002001int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2002 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09002003{
Dennis Zhou (Facebook)b9c39442017-07-24 19:02:01 -04002004 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04002005 size_t static_size, dyn_size;
Dennis Zhou (Facebook)0c4169c2017-07-24 19:02:04 -04002006 struct pcpu_chunk *chunk;
Tejun Heo65632972009-08-14 15:00:52 +09002007 unsigned long *group_offsets;
2008 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09002009 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09002010 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002011 int *unit_map;
2012 int group, unit, i;
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04002013 int map_size;
2014 unsigned long tmp_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09002015
Tejun Heo635b75f2009-09-24 09:43:11 +09002016#define PCPU_SETUP_BUG_ON(cond) do { \
2017 if (unlikely(cond)) { \
Joe Perches870d4b12016-03-17 14:19:53 -07002018 pr_emerg("failed to initialize, %s\n", #cond); \
2019 pr_emerg("cpu_possible_mask=%*pb\n", \
Tejun Heo807de072015-02-13 14:37:34 -08002020 cpumask_pr_args(cpu_possible_mask)); \
Tejun Heo635b75f2009-09-24 09:43:11 +09002021 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2022 BUG(); \
2023 } \
2024} while (0)
2025
Tejun Heo2f39e632009-07-04 08:11:00 +09002026 /* sanity checks */
Tejun Heo635b75f2009-09-24 09:43:11 +09002027 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
Tejun Heobbddff02010-09-03 18:22:48 +02002028#ifdef CONFIG_SMP
Tejun Heo635b75f2009-09-24 09:43:11 +09002029 PCPU_SETUP_BUG_ON(!ai->static_size);
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08002030 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
Tejun Heobbddff02010-09-03 18:22:48 +02002031#endif
Tejun Heo635b75f2009-09-24 09:43:11 +09002032 PCPU_SETUP_BUG_ON(!base_addr);
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08002033 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
Tejun Heo635b75f2009-09-24 09:43:11 +09002034 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08002035 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
Tejun Heo635b75f2009-09-24 09:43:11 +09002036 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04002037 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
Tejun Heo099a19d2010-06-27 18:50:00 +02002038 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
Dennis Zhou (Facebook)fb29a2c2017-07-24 19:01:58 -04002039 PCPU_SETUP_BUG_ON(!ai->dyn_size);
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04002040 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
Dennis Zhou (Facebook)ca460b32017-07-24 19:02:12 -04002041 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2042 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
Tejun Heo9f645532010-04-09 18:57:01 +09002043 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09002044
Tejun Heo65632972009-08-14 15:00:52 +09002045 /* process group information and build config tables accordingly */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002046 group_offsets = memblock_virt_alloc(ai->nr_groups *
2047 sizeof(group_offsets[0]), 0);
2048 group_sizes = memblock_virt_alloc(ai->nr_groups *
2049 sizeof(group_sizes[0]), 0);
2050 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
2051 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
Tejun Heo2f39e632009-07-04 08:11:00 +09002052
Tejun Heofd1e8a12009-08-14 15:00:51 +09002053 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09002054 unit_map[cpu] = UINT_MAX;
Tejun Heoa855b842011-11-18 10:55:35 -08002055
2056 pcpu_low_unit_cpu = NR_CPUS;
2057 pcpu_high_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09002058
Tejun Heofd1e8a12009-08-14 15:00:51 +09002059 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2060 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09002061
Tejun Heo65632972009-08-14 15:00:52 +09002062 group_offsets[group] = gi->base_offset;
2063 group_sizes[group] = gi->nr_units * ai->unit_size;
2064
Tejun Heofd1e8a12009-08-14 15:00:51 +09002065 for (i = 0; i < gi->nr_units; i++) {
2066 cpu = gi->cpu_map[i];
2067 if (cpu == NR_CPUS)
2068 continue;
2069
Dan Carpenter9f295662014-10-29 11:45:04 +03002070 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
Tejun Heo635b75f2009-09-24 09:43:11 +09002071 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2072 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002073
2074 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09002075 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2076
Tejun Heoa855b842011-11-18 10:55:35 -08002077 /* determine low/high unit_cpu */
2078 if (pcpu_low_unit_cpu == NR_CPUS ||
2079 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2080 pcpu_low_unit_cpu = cpu;
2081 if (pcpu_high_unit_cpu == NR_CPUS ||
2082 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2083 pcpu_high_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09002084 }
Tejun Heo2f39e632009-07-04 08:11:00 +09002085 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09002086 pcpu_nr_units = unit;
2087
2088 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09002089 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2090
2091 /* we're done parsing the input, undefine BUG macro and dump config */
2092#undef PCPU_SETUP_BUG_ON
Tejun Heobcbea792010-12-22 14:19:14 +01002093 pcpu_dump_alloc_info(KERN_DEBUG, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002094
Tejun Heo65632972009-08-14 15:00:52 +09002095 pcpu_nr_groups = ai->nr_groups;
2096 pcpu_group_offsets = group_offsets;
2097 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002098 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09002099 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09002100
2101 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002102 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09002103 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09002104 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09002105 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2106 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09002107
Dennis Zhou30a5b532017-06-19 19:28:31 -04002108 pcpu_stats_save_ai(ai);
2109
Tejun Heod9b55ee2009-02-24 11:57:21 +09002110 /*
2111 * Allocate chunk slots. The additional last slot is for
2112 * empty chunks.
2113 */
2114 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002115 pcpu_slot = memblock_virt_alloc(
2116 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09002117 for (i = 0; i < pcpu_nr_slots; i++)
2118 INIT_LIST_HEAD(&pcpu_slot[i]);
2119
Tejun Heoedcb4632009-03-06 14:33:59 +09002120 /*
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04002121 * The end of the static region needs to be aligned with the
2122 * minimum allocation size as this offsets the reserved and
2123 * dynamic region. The first chunk ends page aligned by
2124 * expanding the dynamic region, therefore the dynamic region
2125 * can be shrunk to compensate while still staying above the
2126 * configured sizes.
2127 */
2128 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2129 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2130
2131 /*
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04002132 * Initialize first chunk.
2133 * If the reserved_size is non-zero, this initializes the reserved
2134 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2135 * and the dynamic region is initialized here. The first chunk,
2136 * pcpu_first_chunk, will always point to the chunk that serves
2137 * the dynamic region.
Tejun Heoedcb4632009-03-06 14:33:59 +09002138 */
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04002139 tmp_addr = (unsigned long)base_addr + static_size;
2140 map_size = ai->reserved_size ?: dyn_size;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07002141 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
Tejun Heo61ace7f2009-03-06 14:33:59 +09002142
Tejun Heoedcb4632009-03-06 14:33:59 +09002143 /* init dynamic chunk if necessary */
Dennis Zhou (Facebook)b9c39442017-07-24 19:02:01 -04002144 if (ai->reserved_size) {
Dennis Zhou (Facebook)0c4169c2017-07-24 19:02:04 -04002145 pcpu_reserved_chunk = chunk;
Dennis Zhou (Facebook)b9c39442017-07-24 19:02:01 -04002146
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04002147 tmp_addr = (unsigned long)base_addr + static_size +
Dennis Zhou (Facebook)c0ebfdc2017-07-24 19:02:05 -04002148 ai->reserved_size;
Dennis Zhou (Facebook)d2f3c382017-07-24 19:02:09 -04002149 map_size = dyn_size;
Dennis Zhou (Facebook)40064ae2017-07-12 11:27:32 -07002150 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
Tejun Heoedcb4632009-03-06 14:33:59 +09002151 }
2152
Tejun Heo2441d152009-03-06 14:33:59 +09002153 /* link the first chunk in */
Dennis Zhou (Facebook)0c4169c2017-07-24 19:02:04 -04002154 pcpu_first_chunk = chunk;
Dennis Zhou (Facebook)0cecf502017-07-24 19:02:08 -04002155 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09002156 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09002157
Dennis Zhou30a5b532017-06-19 19:28:31 -04002158 pcpu_stats_chunk_alloc();
Dennis Zhoudf95e792017-06-19 19:28:32 -04002159 trace_percpu_create_chunk(base_addr);
Dennis Zhou30a5b532017-06-19 19:28:31 -04002160
Tejun Heofbf59bc2009-02-20 16:29:08 +09002161 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09002162 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09002163 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09002164}
Tejun Heo66c3a752009-03-10 16:27:48 +09002165
Tejun Heobbddff02010-09-03 18:22:48 +02002166#ifdef CONFIG_SMP
2167
Andi Kleen17f36092012-10-04 17:12:07 -07002168const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
Tejun Heof58dc012009-08-14 15:00:50 +09002169 [PCPU_FC_AUTO] = "auto",
2170 [PCPU_FC_EMBED] = "embed",
2171 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09002172};
Tejun Heo66c3a752009-03-10 16:27:48 +09002173
Tejun Heof58dc012009-08-14 15:00:50 +09002174enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2175
2176static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09002177{
Cyrill Gorcunov5479c782012-11-25 01:17:13 +04002178 if (!str)
2179 return -EINVAL;
2180
Tejun Heof58dc012009-08-14 15:00:50 +09002181 if (0)
2182 /* nada */;
2183#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2184 else if (!strcmp(str, "embed"))
2185 pcpu_chosen_fc = PCPU_FC_EMBED;
2186#endif
2187#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2188 else if (!strcmp(str, "page"))
2189 pcpu_chosen_fc = PCPU_FC_PAGE;
2190#endif
Tejun Heof58dc012009-08-14 15:00:50 +09002191 else
Joe Perches870d4b12016-03-17 14:19:53 -07002192 pr_warn("unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09002193
Tejun Heof58dc012009-08-14 15:00:50 +09002194 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09002195}
Tejun Heof58dc012009-08-14 15:00:50 +09002196early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09002197
Tejun Heo3c9a0242010-09-09 18:00:15 +02002198/*
2199 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2200 * Build it if needed by the arch config or the generic setup is going
2201 * to be used.
2202 */
Tejun Heo08fc4582009-08-14 15:00:49 +09002203#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2204 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo3c9a0242010-09-09 18:00:15 +02002205#define BUILD_EMBED_FIRST_CHUNK
2206#endif
2207
2208/* build pcpu_page_first_chunk() iff needed by the arch config */
2209#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2210#define BUILD_PAGE_FIRST_CHUNK
2211#endif
2212
2213/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2214#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2215/**
Tejun Heofbf59bc2009-02-20 16:29:08 +09002216 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2217 * @reserved_size: the size of reserved percpu area in bytes
2218 * @dyn_size: minimum free size for dynamic allocation in bytes
2219 * @atom_size: allocation atom size
2220 * @cpu_distance_fn: callback to determine distance between cpus, optional
2221 *
2222 * This function determines grouping of units, their mappings to cpus
2223 * and other parameters considering needed percpu size, allocation
2224 * atom size and distances between CPUs.
2225 *
Yannick Guerrinibffc4372015-03-06 23:30:42 +01002226 * Groups are always multiples of atom size and CPUs which are of
Tejun Heofbf59bc2009-02-20 16:29:08 +09002227 * LOCAL_DISTANCE both ways are grouped together and share space for
2228 * units in the same group. The returned configuration is guaranteed
2229 * to have CPUs on different nodes on different groups and >=75% usage
2230 * of allocated virtual address space.
2231 *
2232 * RETURNS:
2233 * On success, pointer to the new allocation_info is returned. On
2234 * failure, ERR_PTR value is returned.
2235 */
2236static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2237 size_t reserved_size, size_t dyn_size,
2238 size_t atom_size,
2239 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2240{
2241 static int group_map[NR_CPUS] __initdata;
2242 static int group_cnt[NR_CPUS] __initdata;
2243 const size_t static_size = __per_cpu_end - __per_cpu_start;
2244 int nr_groups = 1, nr_units = 0;
2245 size_t size_sum, min_unit_size, alloc_size;
2246 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
2247 int last_allocs, group, unit;
2248 unsigned int cpu, tcpu;
2249 struct pcpu_alloc_info *ai;
2250 unsigned int *cpu_map;
2251
2252 /* this function may be called multiple times */
2253 memset(group_map, 0, sizeof(group_map));
2254 memset(group_cnt, 0, sizeof(group_cnt));
2255
2256 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2257 size_sum = PFN_ALIGN(static_size + reserved_size +
2258 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2259 dyn_size = size_sum - static_size - reserved_size;
2260
2261 /*
2262 * Determine min_unit_size, alloc_size and max_upa such that
2263 * alloc_size is multiple of atom_size and is the smallest
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002264 * which can accommodate 4k aligned segments which are equal to
Tejun Heofbf59bc2009-02-20 16:29:08 +09002265 * or larger than min_unit_size.
2266 */
2267 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2268
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -04002269 /* determine the maximum # of units that can fit in an allocation */
Tejun Heofbf59bc2009-02-20 16:29:08 +09002270 alloc_size = roundup(min_unit_size, atom_size);
2271 upa = alloc_size / min_unit_size;
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08002272 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
Tejun Heofbf59bc2009-02-20 16:29:08 +09002273 upa--;
2274 max_upa = upa;
2275
2276 /* group cpus according to their proximity */
2277 for_each_possible_cpu(cpu) {
2278 group = 0;
2279 next_group:
2280 for_each_possible_cpu(tcpu) {
2281 if (cpu == tcpu)
2282 break;
2283 if (group_map[tcpu] == group && cpu_distance_fn &&
2284 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2285 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2286 group++;
2287 nr_groups = max(nr_groups, group + 1);
2288 goto next_group;
2289 }
2290 }
2291 group_map[cpu] = group;
2292 group_cnt[group]++;
2293 }
2294
2295 /*
Dennis Zhou (Facebook)9c015162017-07-15 22:23:09 -04002296 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2297 * Expand the unit_size until we use >= 75% of the units allocated.
2298 * Related to atom_size, which could be much larger than the unit_size.
Tejun Heofbf59bc2009-02-20 16:29:08 +09002299 */
2300 last_allocs = INT_MAX;
2301 for (upa = max_upa; upa; upa--) {
2302 int allocs = 0, wasted = 0;
2303
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08002304 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
Tejun Heofbf59bc2009-02-20 16:29:08 +09002305 continue;
2306
2307 for (group = 0; group < nr_groups; group++) {
2308 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2309 allocs += this_allocs;
2310 wasted += this_allocs * upa - group_cnt[group];
2311 }
2312
2313 /*
2314 * Don't accept if wastage is over 1/3. The
2315 * greater-than comparison ensures upa==1 always
2316 * passes the following check.
2317 */
2318 if (wasted > num_possible_cpus() / 3)
2319 continue;
2320
2321 /* and then don't consume more memory */
2322 if (allocs > last_allocs)
2323 break;
2324 last_allocs = allocs;
2325 best_upa = upa;
2326 }
2327 upa = best_upa;
2328
2329 /* allocate and fill alloc_info */
2330 for (group = 0; group < nr_groups; group++)
2331 nr_units += roundup(group_cnt[group], upa);
2332
2333 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2334 if (!ai)
2335 return ERR_PTR(-ENOMEM);
2336 cpu_map = ai->groups[0].cpu_map;
2337
2338 for (group = 0; group < nr_groups; group++) {
2339 ai->groups[group].cpu_map = cpu_map;
2340 cpu_map += roundup(group_cnt[group], upa);
2341 }
2342
2343 ai->static_size = static_size;
2344 ai->reserved_size = reserved_size;
2345 ai->dyn_size = dyn_size;
2346 ai->unit_size = alloc_size / upa;
2347 ai->atom_size = atom_size;
2348 ai->alloc_size = alloc_size;
2349
2350 for (group = 0, unit = 0; group_cnt[group]; group++) {
2351 struct pcpu_group_info *gi = &ai->groups[group];
2352
2353 /*
2354 * Initialize base_offset as if all groups are located
2355 * back-to-back. The caller should update this to
2356 * reflect actual allocation.
2357 */
2358 gi->base_offset = unit * ai->unit_size;
2359
2360 for_each_possible_cpu(cpu)
2361 if (group_map[cpu] == group)
2362 gi->cpu_map[gi->nr_units++] = cpu;
2363 gi->nr_units = roundup(gi->nr_units, upa);
2364 unit += gi->nr_units;
2365 }
2366 BUG_ON(unit != nr_units);
2367
2368 return ai;
2369}
Tejun Heo3c9a0242010-09-09 18:00:15 +02002370#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
Tejun Heofbf59bc2009-02-20 16:29:08 +09002371
Tejun Heo3c9a0242010-09-09 18:00:15 +02002372#if defined(BUILD_EMBED_FIRST_CHUNK)
Tejun Heo66c3a752009-03-10 16:27:48 +09002373/**
2374 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09002375 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02002376 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heoc8826dd2009-08-14 15:00:52 +09002377 * @atom_size: allocation atom size
2378 * @cpu_distance_fn: callback to determine distance between cpus, optional
2379 * @alloc_fn: function to allocate percpu page
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002380 * @free_fn: function to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09002381 *
2382 * This is a helper to ease setting up embedded first percpu chunk and
2383 * can be called where pcpu_setup_first_chunk() is expected.
2384 *
2385 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09002386 * by calling @alloc_fn and used as-is without being mapped into
2387 * vmalloc area. Allocations are always whole multiples of @atom_size
2388 * aligned to @atom_size.
2389 *
2390 * This enables the first chunk to piggy back on the linear physical
2391 * mapping which often uses larger page size. Please note that this
2392 * can result in very sparse cpu->unit mapping on NUMA machines thus
2393 * requiring large vmalloc address space. Don't use this allocator if
2394 * vmalloc space is not orders of magnitude larger than distances
2395 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09002396 *
Tejun Heo4ba6ce22010-06-27 18:49:59 +02002397 * @dyn_size specifies the minimum dynamic area size.
Tejun Heo66c3a752009-03-10 16:27:48 +09002398 *
2399 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09002400 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09002401 *
2402 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09002403 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09002404 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02002405int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
Tejun Heoc8826dd2009-08-14 15:00:52 +09002406 size_t atom_size,
2407 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2408 pcpu_fc_alloc_fn_t alloc_fn,
2409 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09002410{
Tejun Heoc8826dd2009-08-14 15:00:52 +09002411 void *base = (void *)ULONG_MAX;
2412 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002413 struct pcpu_alloc_info *ai;
zijun_hu93c76b6b2016-10-05 21:19:11 +08002414 size_t size_sum, areas_size;
2415 unsigned long max_distance;
zijun_hu9b739662016-10-05 21:30:24 +08002416 int group, i, highest_group, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09002417
Tejun Heoc8826dd2009-08-14 15:00:52 +09002418 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2419 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002420 if (IS_ERR(ai))
2421 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09002422
Tejun Heofd1e8a12009-08-14 15:00:51 +09002423 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09002424 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09002425
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002426 areas = memblock_virt_alloc_nopanic(areas_size, 0);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002427 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09002428 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09002429 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09002430 }
Tejun Heo66c3a752009-03-10 16:27:48 +09002431
zijun_hu9b739662016-10-05 21:30:24 +08002432 /* allocate, copy and determine base address & max_distance */
2433 highest_group = 0;
Tejun Heoc8826dd2009-08-14 15:00:52 +09002434 for (group = 0; group < ai->nr_groups; group++) {
2435 struct pcpu_group_info *gi = &ai->groups[group];
2436 unsigned int cpu = NR_CPUS;
2437 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09002438
Tejun Heoc8826dd2009-08-14 15:00:52 +09002439 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2440 cpu = gi->cpu_map[i];
2441 BUG_ON(cpu == NR_CPUS);
2442
2443 /* allocate space for the whole group */
2444 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2445 if (!ptr) {
2446 rc = -ENOMEM;
2447 goto out_free_areas;
2448 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01002449 /* kmemleak tracks the percpu allocations separately */
2450 kmemleak_free(ptr);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002451 areas[group] = ptr;
2452
2453 base = min(ptr, base);
zijun_hu9b739662016-10-05 21:30:24 +08002454 if (ptr > areas[highest_group])
2455 highest_group = group;
2456 }
2457 max_distance = areas[highest_group] - base;
2458 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2459
2460 /* warn if maximum distance is further than 75% of vmalloc space */
2461 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2462 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2463 max_distance, VMALLOC_TOTAL);
2464#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2465 /* and fail if we have fallback */
2466 rc = -EINVAL;
2467 goto out_free_areas;
2468#endif
Tejun Heo42b64282012-04-27 08:42:53 -07002469 }
2470
2471 /*
2472 * Copy data and free unused parts. This should happen after all
2473 * allocations are complete; otherwise, we may end up with
2474 * overlapping groups.
2475 */
2476 for (group = 0; group < ai->nr_groups; group++) {
2477 struct pcpu_group_info *gi = &ai->groups[group];
2478 void *ptr = areas[group];
Tejun Heoc8826dd2009-08-14 15:00:52 +09002479
2480 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2481 if (gi->cpu_map[i] == NR_CPUS) {
2482 /* unused unit, free whole */
2483 free_fn(ptr, ai->unit_size);
2484 continue;
2485 }
2486 /* copy and return the unused part */
2487 memcpy(ptr, __per_cpu_load, ai->static_size);
2488 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2489 }
Tejun Heo66c3a752009-03-10 16:27:48 +09002490 }
2491
Tejun Heoc8826dd2009-08-14 15:00:52 +09002492 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09002493 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09002494 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo6ea529a2009-09-24 18:46:01 +09002495 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09002496
Joe Perches870d4b12016-03-17 14:19:53 -07002497 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002498 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2499 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002500
Tejun Heofb435d52009-08-14 15:00:51 +09002501 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002502 goto out_free;
2503
2504out_free_areas:
2505 for (group = 0; group < ai->nr_groups; group++)
Michael Holzheuf851c8d2013-09-17 16:57:34 +02002506 if (areas[group])
2507 free_fn(areas[group],
2508 ai->groups[group].nr_units * ai->unit_size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002509out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09002510 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002511 if (areas)
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002512 memblock_free_early(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09002513 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09002514}
Tejun Heo3c9a0242010-09-09 18:00:15 +02002515#endif /* BUILD_EMBED_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002516
Tejun Heo3c9a0242010-09-09 18:00:15 +02002517#ifdef BUILD_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09002518/**
Tejun Heo00ae4062009-08-14 15:00:49 +09002519 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09002520 * @reserved_size: the size of reserved percpu area in bytes
2521 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002522 * @free_fn: function to free percpu page, always called with PAGE_SIZE
Tejun Heod4b95f82009-07-04 08:10:59 +09002523 * @populate_pte_fn: function to populate pte
2524 *
Tejun Heo00ae4062009-08-14 15:00:49 +09002525 * This is a helper to ease setting up page-remapped first percpu
2526 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09002527 *
2528 * This is the basic allocator. Static percpu area is allocated
2529 * page-by-page into vmalloc area.
2530 *
2531 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09002532 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09002533 */
Tejun Heofb435d52009-08-14 15:00:51 +09002534int __init pcpu_page_first_chunk(size_t reserved_size,
2535 pcpu_fc_alloc_fn_t alloc_fn,
2536 pcpu_fc_free_fn_t free_fn,
2537 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09002538{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002539 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002540 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09002541 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09002542 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09002543 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09002544 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09002545 int unit, i, j, rc;
zijun_hu8f606602016-12-12 16:45:02 -08002546 int upa;
2547 int nr_g0_units;
Tejun Heod4b95f82009-07-04 08:10:59 +09002548
Tejun Heo00ae4062009-08-14 15:00:49 +09002549 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2550
Tejun Heo4ba6ce22010-06-27 18:49:59 +02002551 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002552 if (IS_ERR(ai))
2553 return PTR_ERR(ai);
2554 BUG_ON(ai->nr_groups != 1);
zijun_hu8f606602016-12-12 16:45:02 -08002555 upa = ai->alloc_size/ai->unit_size;
2556 nr_g0_units = roundup(num_possible_cpus(), upa);
2557 if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) {
2558 pcpu_free_alloc_info(ai);
2559 return -EINVAL;
2560 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09002561
2562 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09002563
2564 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002565 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2566 sizeof(pages[0]));
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002567 pages = memblock_virt_alloc(pages_size, 0);
Tejun Heod4b95f82009-07-04 08:10:59 +09002568
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002569 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09002570 j = 0;
zijun_hu8f606602016-12-12 16:45:02 -08002571 for (unit = 0; unit < num_possible_cpus(); unit++) {
2572 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heoce3141a2009-07-04 08:11:00 +09002573 for (i = 0; i < unit_pages; i++) {
Tejun Heod4b95f82009-07-04 08:10:59 +09002574 void *ptr;
2575
Tejun Heo3cbc8562009-08-14 15:00:50 +09002576 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09002577 if (!ptr) {
Joe Perches870d4b12016-03-17 14:19:53 -07002578 pr_warn("failed to allocate %s page for cpu%u\n",
zijun_hu8f606602016-12-12 16:45:02 -08002579 psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09002580 goto enomem;
2581 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01002582 /* kmemleak tracks the percpu allocations separately */
2583 kmemleak_free(ptr);
Tejun Heoce3141a2009-07-04 08:11:00 +09002584 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002585 }
zijun_hu8f606602016-12-12 16:45:02 -08002586 }
Tejun Heod4b95f82009-07-04 08:10:59 +09002587
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002588 /* allocate vm area, map the pages and copy static data */
2589 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002590 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002591 vm_area_register_early(&vm, PAGE_SIZE);
2592
Tejun Heofd1e8a12009-08-14 15:00:51 +09002593 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09002594 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09002595 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002596
Tejun Heoce3141a2009-07-04 08:11:00 +09002597 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002598 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2599
2600 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09002601 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2602 unit_pages);
2603 if (rc < 0)
2604 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002605
2606 /*
2607 * FIXME: Archs with virtual cache should flush local
2608 * cache for the linear mapping here - something
2609 * equivalent to flush_cache_vmap() on the local cpu.
2610 * flush_cache_vmap() can't be used as most supporting
2611 * data structures are not set up yet.
2612 */
2613
2614 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002615 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002616 }
2617
2618 /* we're ready, commit */
Joe Perches870d4b12016-03-17 14:19:53 -07002619 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002620 unit_pages, psize_str, vm.addr, ai->static_size,
2621 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002622
Tejun Heofb435d52009-08-14 15:00:51 +09002623 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002624 goto out_free_ar;
2625
2626enomem:
2627 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09002628 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09002629 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09002630out_free_ar:
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002631 memblock_free_early(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002632 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09002633 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09002634}
Tejun Heo3c9a0242010-09-09 18:00:15 +02002635#endif /* BUILD_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002636
Tejun Heobbddff02010-09-03 18:22:48 +02002637#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
Tejun Heo8c4bfc62009-07-04 08:10:59 +09002638/*
Tejun Heobbddff02010-09-03 18:22:48 +02002639 * Generic SMP percpu area setup.
Tejun Heoe74e3962009-03-30 19:07:44 +09002640 *
2641 * The embedding helper is used because its behavior closely resembles
2642 * the original non-dynamic generic percpu area setup. This is
2643 * important because many archs have addressing restrictions and might
2644 * fail if the percpu area is located far away from the previous
2645 * location. As an added bonus, in non-NUMA cases, embedding is
2646 * generally a good idea TLB-wise because percpu area can piggy back
2647 * on the physical linear memory mapping which uses large page
2648 * mappings on applicable archs.
2649 */
Tejun Heoe74e3962009-03-30 19:07:44 +09002650unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2651EXPORT_SYMBOL(__per_cpu_offset);
2652
Tejun Heoc8826dd2009-08-14 15:00:52 +09002653static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2654 size_t align)
2655{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002656 return memblock_virt_alloc_from_nopanic(
2657 size, align, __pa(MAX_DMA_ADDRESS));
Tejun Heoc8826dd2009-08-14 15:00:52 +09002658}
2659
2660static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2661{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002662 memblock_free_early(__pa(ptr), size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002663}
2664
Tejun Heoe74e3962009-03-30 19:07:44 +09002665void __init setup_per_cpu_areas(void)
2666{
Tejun Heoe74e3962009-03-30 19:07:44 +09002667 unsigned long delta;
2668 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09002669 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09002670
2671 /*
2672 * Always reserve area for module percpu variables. That's
2673 * what the legacy allocator did.
2674 */
Tejun Heofb435d52009-08-14 15:00:51 +09002675 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09002676 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2677 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09002678 if (rc < 0)
Tejun Heobbddff02010-09-03 18:22:48 +02002679 panic("Failed to initialize percpu areas.");
Tejun Heoe74e3962009-03-30 19:07:44 +09002680
2681 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2682 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09002683 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09002684}
Tejun Heobbddff02010-09-03 18:22:48 +02002685#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2686
2687#else /* CONFIG_SMP */
2688
2689/*
2690 * UP percpu area setup.
2691 *
2692 * UP always uses km-based percpu allocator with identity mapping.
2693 * Static percpu variables are indistinguishable from the usual static
2694 * variables and don't require any special preparation.
2695 */
2696void __init setup_per_cpu_areas(void)
2697{
2698 const size_t unit_size =
2699 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2700 PERCPU_DYNAMIC_RESERVE));
2701 struct pcpu_alloc_info *ai;
2702 void *fc;
2703
2704 ai = pcpu_alloc_alloc_info(1, 1);
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002705 fc = memblock_virt_alloc_from_nopanic(unit_size,
2706 PAGE_SIZE,
2707 __pa(MAX_DMA_ADDRESS));
Tejun Heobbddff02010-09-03 18:22:48 +02002708 if (!ai || !fc)
2709 panic("Failed to allocate memory for percpu areas.");
Catalin Marinas100d13c2012-05-09 16:55:19 +01002710 /* kmemleak tracks the percpu allocations separately */
2711 kmemleak_free(fc);
Tejun Heobbddff02010-09-03 18:22:48 +02002712
2713 ai->dyn_size = unit_size;
2714 ai->unit_size = unit_size;
2715 ai->atom_size = unit_size;
2716 ai->alloc_size = unit_size;
2717 ai->groups[0].nr_units = 1;
2718 ai->groups[0].cpu_map[0] = 0;
2719
2720 if (pcpu_setup_first_chunk(ai, fc) < 0)
2721 panic("Failed to initialize percpu areas.");
2722}
2723
2724#endif /* CONFIG_SMP */
Tejun Heo099a19d2010-06-27 18:50:00 +02002725
2726/*
Tejun Heo1a4d7602014-09-02 14:46:05 -04002727 * Percpu allocator is initialized early during boot when neither slab or
2728 * workqueue is available. Plug async management until everything is up
2729 * and running.
2730 */
2731static int __init percpu_enable_async(void)
2732{
2733 pcpu_async_enabled = true;
2734 return 0;
2735}
2736subsys_initcall(percpu_enable_async);